doi
stringlengths 10
10
| chunk-id
int64 0
936
| chunk
stringlengths 401
2.02k
| id
stringlengths 12
14
| title
stringlengths 8
162
| summary
stringlengths 228
1.92k
| source
stringlengths 31
31
| authors
stringlengths 7
6.97k
| categories
stringlengths 5
107
| comment
stringlengths 4
398
⌀ | journal_ref
stringlengths 8
194
⌀ | primary_category
stringlengths 5
17
| published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1706.06064 | 38 | Different from the above approaches, Cao et al. modeled the global spatial context by two families of ordered bag- of-features as a generation of the spatial pyramid match- ing [128] by linear projection and circular projection and further reï¬ned them to capture the invariance of object translation, rotation, and scaling by simple histogram op- erations, including calibration, equalization, and decompo- sition [129].
In the scenario of face retrieval, the above general code- book generation methods are likely to fail to capture the unique facial characteristics. To generate discriminative vi- sual codebook, Wu et al. proposed to generate identity-based visual vocabulary with some training persons each with multiple face examples under various poses, expressions, and illumination conditions [130]. A visual word is deï¬ned as a tuple consisting of two components, i.e., person ID and position ID and associated with multiple examples.
# 4.4 Feature Quantization
With visual codebook deï¬ned, feature quantization is to assign a visual word ID to each feature. To design a suitable assignment function, special consideration should be made to balance quantization accuracy, efï¬ciency, and memory overhead. | 1706.06064#38 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 39 | The most naive choice is to take the nearest neighbor search, so as to ï¬nd the closest (the most similar) visual word of a given feature by linear scan, which, however, suffers expensive computational cost. Usually, approximate nearest neighbor (ANN) search methods are adopted to speed up the searching process, with sacriï¬ce of accuracy to some extent. In [8], a k-d tree structure [131] is utilized with a best-bin-ï¬rst modiï¬cation to ï¬nd approximate nearest neighbors to the descriptor vector of the query. In [10], based on the hierarchical vocabulary tree, an efï¬cient approximate | 1706.06064#39 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 40 | nearest neighbor search is achieved by propagating the query feature vector from the root node down the tree by comparing the corresponding child nodes and choosing the closest one. In [132], a k-d forest approximation algorithm is proposed with reduced time complexity. Muja and Lowe proposed a novel priority search k-means tree algorithm for scalable nearest neighbor search [133] with FLANN library8 provided. In [118], the feature quantization is achieved by range-based neighbor search over the random seeding code- book. This random seeding approach, although efï¬cient in implementation, suffers the bias to the training data and achieves limited retrieval accuracy in large-scale image retrieval [134]. Those approaches conduct quantization in a hard manner and inevitably incur severe quantization loss. Considering that the codebook partitions the feature space into some non-overlapping cells, feature quantization works to identify which cell a test feature falls into. When the codebook size is large which means the feature space is ï¬nely partitioned, features proximate to the partition boundary are likely to fall into different cells. On the other hand, with small codebook and feature space coarsely par- titioned, | 1706.06064#40 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 41 | features proximate to the partition boundary are likely to fall into different cells. On the other hand, with small codebook and feature space coarsely par- titioned, irrelevant features with large distance may also fall into the same cell. Both cases will incur quantization loss and degrade the recall and precision of feature matching, respectively. A trade-off shall be made on the codebook size to balance the recall and precision from the above two kinds of loss [10], or some constraints are involved to reï¬ne the quantization quality. | 1706.06064#41 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 42 | Some approaches adopt a large visual codebook but take account of the soft quantization to reduce the quantiza- tion loss. Generally, a descriptor-dependent soft assignment scheme [15] is used to map a feature vector to a weighted combination of multiple visual words. Intuitively, the soft quantization can be performed for both a query feature and the database features. However, it will cost several times more memory to store the multiple quantization results for each database feature. As a trade-off, the soft quantization can be constrained to only the query side [35]. In [35], a new quantizer is designed based on a codebook learned by brute-force k-means clustering. It ï¬rst performs k-means clustering on the pre-trained visual words and generate a two-layer visual vocabulary tree in a bottom-up way. Then, new connections between the two-layer nodes are constructed by quantizing a large feature set with both layers of quantizers. Soft assignment is performed with a criteria based on distance ratio. | 1706.06064#42 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 43 | On the other hand, some other approaches keep a rela- tively small visual codebook but performs further veriï¬ca- tion to reduce the quantization loss. In [12], Hamming Em- bedding reduces the dimension of SIFT descriptors quan- tized to a visual word, and then trains a median vector by taking the median value in each dimension of the feature samples. After a new feature is quantized to a visual word, it is projected to the low dimensional space, and then com- pared with the median vector dimension-wise to generate binary signature for matching veriï¬cation [54]. In [135], a variant, i.e., the asymmetric Hamming Embedding scheme, is proposed to exploit the rich information conveyed by the binary signature. Zhou et al.adopt a similar veriï¬cation
8. http://www.cs.ubc.ca/research/ï¬ann/
7
idea with a different binary signature which is generated by comparing each element of a feature descriptor to its median [136]. | 1706.06064#43 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 44 | 7
idea with a different binary signature which is generated by comparing each element of a feature descriptor to its median [136].
The above approaches rely on single visual codebook for feature quantization. To correct quantization artifacts and improve recall, typically, multiple vocabularies are generated for feature quantization to improve the re- call [137][138]. Since multiple vocabularies suffers from vocabulary correlation, Zheng et al proposed a Bayes merg- ing approach to down-weight the indexed features in the intersection set of multiple vocabularies [139]. It models the the correlation problem in a probabilistic view and estimate a joint similarity on both image- and feature-level for the indexed features in the intersection set. | 1706.06064#44 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 45 | The vector quantization of local descriptors is closely related to approximate nearest neighbor search [58]. there are many hashing algorithms for In literature, approximate nearest neighbor (ANN) search, such as LSH [140][141], multi-probe LSH [142], kernelized LSH [56], semi-supervised hashing method (SSH) [143], spectral hash- iterative quantization [144], ing [57], min-Hashing [16], random grids [145], bucket distance hashing (BDH) [146], query-driven iterated neighborhood graph search [147], and linear distance preserving hashing [148]. These hashing methods, however, are mostly applied to global image fea- tures such as GIST or BoW features at the image level, or to feature retrieval only at the local feature level. There is few work dedicated to image level search based on local feature hashing [22]. The major concern of those hashing methods is that multiple hashing tables are usually involved and each feature needs to be indexed multiple times, which cast heavy memory burden. Besides, in hashing methods such as LSH [141], multi-probe LSH [142] and kernelized LSH [56], the original database feature vectors need be | 1706.06064#45 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 46 | Besides, in hashing methods such as LSH [141], multi-probe LSH [142] and kernelized LSH [56], the original database feature vectors need be kept in memory to compute the exact distance to the query feature, which is infeasible in the scenario of large-scale image search with local features. Moreover, approximate nearest neighbor search usually targets at identifying the top-k closest data to the query, which ignores the essence of range-based neighbor search in visual feature matching. That is, given a query feature, the number of target data in the database is query-sensitive and determined by the coverage of the range-based neighborhood of the query. | 1706.06064#46 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 47 | In [58], a novel product quantization is proposed to generate an exponentially large codebook with low cost in memory and time for approximate nearest neighbor search. It decomposes the feature space into a Cartesian product of low-dimensional subspaces and quantizes each sub-space individually. The quantization indices of each sub-space are presented as a short code, based on which the Euclidean distance between two feature vectors can be efï¬ciently es- timated by looking up a pre-computed table. The product quantization, however, suffers from exhaustive search for identifying target features, which is prohibitive in large- scale image search [58]. As a partial solution to this bottle neck, vector quantization by k-means can be involved to narrow the search scope and allow the product to focus on a small fraction of indexed features [58]. In [149], the product quantization is optimized with respect to the vector space decomposition and the quantization codebook with two solutions from the non-parametric and the parametric | 1706.06064#47 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 48 | perspectives. Zhou et al. formulated the feature matching as an Ç«âneighborhood problem and approximated it with a dual-resolution quantization scheme for efï¬cient indexing and querying [134]. It performs scalar quantization in coarse and ï¬ne resolutions on each dimension of the data, and cascades the quantization results over all dimensions. The cascaded quantization results in coarse resolution are used to build the index, while the cascaded quantization results in ï¬ne resolutions are transformed to a binary signature for matching veriï¬cation.
In [150], the high dimensional SIFT descriptor space is partitioned into regular lattices. Although demonstrated to work well in image classiï¬cation, in [15], regular lattice quantization is revealed to work much worse than [10] [15] in large scale image search application.
# 4.5 Feature Aggregation | 1706.06064#48 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 49 | # 4.5 Feature Aggregation
When an image is represented by a set of local features, it is necessary to aggregate those local features into a ï¬xed- length vector representation for convenience of similarity comparison between query and database images. Generally, there are three alternatives to achieve this goal. The ï¬rst one is the classic Bag-of-Visual-Words representation, which quantizes each local feature to the closest visual word of a pre-trained visual codebook. The quantization result of a single local feature can be regarded as a high-dimensional binary vector, where the non-zero dimension corresponds to the quantized visual word. By pooling the quantization results of all local features in an image, we obtain a BoW vector with the dimension size as the visual codebook size. In this scheme, the involved visual codebook is usually very large in size and the generated BoW vector is very sparse, which facilitates the use of the inverted ï¬le indexing. | 1706.06064#49 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 50 | The second popular feature aggregation method is the VLAD (vector of locally aggregated descriptors) ap- proach [116], which adopts k-means based vector quantiza- tion and accumulates the quantization residues for features quantized to each visual word and concatenate those accu- mulated vectors into a single vector representation. With compact size, the VLAD vector inherits some important properties from SIFT feature, including invariance to trans- lation, rotation, and scaling. In [151], the VLAD approach is improved by a new intra-normalization scheme and multiple spatial VLAD representation. An in-depth analysis on VLAD is conducted in [152]. In [153], an extension of VLAD is proposed with triangulation embedding scheme and democratic aggregation technique. Further, Tolias et al. encompassed the VLAD vector with various matching schemes [30]. To reduce the computational complexity of the democratic aggregation scheme, Gao et al. proposed a fast scheme with comparable retrieval accuracy perfor- mance [154]. In [155], sparse coding is adopted to encode the local feature descriptors into sparse vectors, which are further aggregated with a | 1706.06064#50 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 51 | mance [154]. In [155], sparse coding is adopted to encode the local feature descriptors into sparse vectors, which are further aggregated with a max-pooling strategy. Liu et al. proposed a hierarchical scheme to build the VLAD vec- tor with SIFT feature [156]. By involving a hidden-layer vocabulary, the distribution of the residue vectors to be aggregated becomes much more uniform, leading to better discrimination for the representation. | 1706.06064#51 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 52 | representation is achieved by global aggregation of all local features in an
8
image, the original VLAD vector sacriï¬ces the ï¬exibility to address partial occlusion and background clutter. To allevi- ate this problem, Liu et al. [157] grouped local key points by their spatial positions in the image plane and aggregated all local descriptors in each group by the VLAD scheme [116]. As a result, a local aggregation of local features is achieved and promising retrieval accuracy is demonstrated with a tradeoff in memory cost. | 1706.06064#52 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 53 | Besides the BoW representation and the VLAD, another alternative is the Fisher Vector based representation [117] with Fisher kernel [158] [159]. As a generative model, given a set of features for an image, Fisher vector represents them into a ï¬x-sized vector by the gradient of the log-likelihood function with respect to a set of parameter vectors [160]. In [117] [161], Gaussian Mixture Model (GMM) is adopted as a generative model to aggregate the normalized con- catenated gradient vectors of all local descriptors into a uniform Fisher vector with an average pooling scheme. In fact, the Fisher Vector can be regarded as a generalized representation of the BoW representation and VLAD. On one hand, if we keep only the gradient of the log-likelihood function with respect to the weight of GMM, the Fisher Vector degenerates to a soft version of the BoW vector. On the other hand, If we keep only the gradient of the log-likelyhood function with respect to the mean vector of GMM, we can derive the VLAD representation [58]. | 1706.06064#53 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 54 | In either the Fish vector or VLAD representation, the in- volved GMM number or codebook size is relative small and the obtained aggregated vector is no long sparse. As a result, it is unsuitable to apply the inverted ï¬le indexing scheme to index images based on the aggregated results. To address this dilemma, the aggregated vector is dimensionally re- duced and further encoded by product quantization [58] for efï¬cient distance computation.
The above aggregation schemes are based on local hand-crafted feature, such as SIFT feature. Intuitively, such schemes can be directly leveraged to local deep features. Following this idea, Gong et al. [162] extract local CNN features from the local patches sampled regularly at mul- tiple scale levels and pool the CNN features in each scale level with the VLAD scheme [37]. In [163], Babenko et al. interpret the activations from the last convolutional layers of CNNs as local deep features. They reveal that the individual similarity of local deep feature is very discriminative and the simple aggregation with sum pooling over local deep feature yields the best performance. | 1706.06064#54 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 55 | 5 DATABASE INDEXING Image index refers to a database organizing structure to assist for efï¬cient retrieval of the target images. Since the response time is a key issue in retrieval, the signiï¬cance of database indexing is becoming increasingly evident as the scale of image database on the Web explosively grows. Generally, in CBIR, two kinds of indexing techniques are popularly adopted, i.e., inverted ï¬le indexing and hashing based indexing. In the following, we will discuss related retrieval algorithms in each category, respectively.
# 5.1 Inverted File Indexing
Inspired by the success of text search engines, inverted ï¬le indexing [164] has been successfully used for large
scale image search [9] [11] [18] [14] [10] [12] [17] [165]. In essence, the inverted ï¬le structure is a compact column- wise representation of a sparse matrix, where the row and the column denote image and visual word, respectively. In on-line retrieval, only those images sharing common visual words with the query image need to be checked. Therefore, the number of candidate images to be compared is greatly reduced, achieving an efï¬cient response. | 1706.06064#55 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 56 | In the inverted ï¬le structure, each visual word is fol- lowed by an inverted ï¬le list of entries. Each entry stores the ID of the image where the visual word appears, and some other clues for veriï¬cation or similarity measurement. For instance, Hamming Embedding [12] generates a 64-bit Hamming code for each feature to verify the descriptor matching. The geometric clues, such as feature position, scale, and orientation, are also stored in the inverted ï¬le list for geometric consistency veriï¬cation [11] [18] [12] [13]. In [17], Wu et al. recorded the feature orders in horizontal and veriï¬cation direction in each bundled feature located in a MSER region. In [123], 3 spatial statistics, including descriptor density, mean relative log scale, and mean orien- tation difference, are calculated for each feature and stored in the inverted list after quantization. Zheng et al. modeled the correlation between multiple features with a multi-IDF scheme and coupled the binary signatures of those features into the inverted ï¬le to enhances the quality of visual matching [166]. | 1706.06064#56 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 57 | Following the general idea of inverted ï¬le structure, many variants are proposed. In [42], to adapt to the in- verted index structure for sketch-based retrieval, it regularly quantizes the edge pixel in position channel and orientation channel and follows each entry in the edgel dictionary with an inverted lists of related images. In [68], Zheng et al proposed a new coupled Multi-Index (c-MI) framework to fuse complementary features at indexing level. Each dimension of c-MI corresponds to one kind of feature, and the retrieval process votes for images similar in both SIFT and color attribute [85] feature spaces. In [70], the image database is cross-indexed in both the binary SIFT space and the original SIFT space. With such cross-indexing structure, a new searching strategy is designed to ï¬nd target data for effective feature quantization. | 1706.06064#57 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 58 | Some methods try to embed the semantics into the index structure. In [167], Zhang et al proposed a new indexing structure by decomposing a document-like representation of an image into two components, one for dimension reduction and the other for residual information preservation. The decomposition is achieved by either a graphical model or a matrix factorization approach. Then, the similarity between images is transferred to measuring similarities of their com- ponents. In [89], Zhang et al proposed a semantic-aware co- indexing to jointly embed two strong cues, i.e., local SIFT feature and semantic attributes, into the inverted indexes. It exploits 1000 semantic attributes to ï¬lter out isolated images and insert semantically similar images to the initial inverted index set built based on local SIFT features. As a result, the discriminative capability of the indexed features is signiï¬cantly enhanced.
To adapt the product quantization [58] to the inverted in- dex idea, inverted multi-index is proposed to generalize the inverted index idea by replacing the standard quantization
9
within inverted indices with product quantization, so as to speed up the approximate nearest neighbor search. | 1706.06064#58 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 59 | 9
within inverted indices with product quantization, so as to speed up the approximate nearest neighbor search.
To improve the recall rate of inverted indexing algo- rithms, the database images are indexed multiple times with multiple quantizers, such as randomized k-d trees [168] [66]. In [137], a joint inverted indexing algorithm is proposed, which jointly optimizes all codewords in all quantizers and demonstrates considerable improvement over methods with multiple independent quantizers. In [23], this goal is achieved by augmenting the image features for the database images which are estimated to be visible in a homograpy in the augmented images. | 1706.06064#59 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 60 | To speedup the online retrieval process, Zheng et al. proposed a novel Q-Index structure based on the inverted index organization [169]. It deï¬nes an impact score for each indexed local SIFT feature based on TF-IDF, scale, saliency, and quantization ambiguity. Then, based on the impact score, it introduced two complementary strategies, i.e. query pruning and early termination, with the former to discard less important features in the query and the later to partially visit the index lists containing the most important indexed features. The proposed algorithm demonstrates signiï¬cant speed-up for online query with competitive retrieval accu- racy. In [170], Ji et al. considered the scenario of parallelized image retrieval and proposed to distribute visual indexing structure over multiple servers. To reduce the search latency across servers, it formulates the index distribution problem as a learning problem by maximizing the uniformity of assigning the words of a given query to multiple servers.
# 5.2 Hashing Based Indexing | 1706.06064#60 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 61 | When the image representation, for instance GIST feature and VLAD feature, is a dense vector with the majority of the coefï¬cients being non-zero, it is unsuitable to directly apply the inverted ï¬le structure for indexing. To achieve efï¬cient retrieval for relevant results, hashing techniques are popularly adopted [171] [172] [173] [174] [175]. The most representative hashing scheme is the locality sensitive hashing (LSH) [176], which partitions the feature space with multiple hash functions of random projections with the intuition that for objects which are close to each other, the collision probability is much higher than for those which are far away. Given a query, some candidates are ï¬rst retrieved based on hashing collision and re-ranked based on the exact distance from the query. In [56], LSH is generated to accommodate arbitrary kernel functions, with sub-linear time approximate similarity search permitted. The potential concern of those hashing scheme is that, since the raw database representation vectors should be stored in memory for the reranking stage, they are not well scalable to large- scale image database. In [177], a feature map is proposed by integrating appearance | 1706.06064#61 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 62 | be stored in memory for the reranking stage, they are not well scalable to large- scale image database. In [177], a feature map is proposed by integrating appearance and global geometry, which is further hashed for indexing. This scheme, however, suffers expensive memory cost which is quadratic in the number of local features, which limits its scalability towards large scale image retrieval. To address this drawback, an extension is made with a feature selection model to replace the hashing approach [178]. | 1706.06064#62 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 63 | With the inverted index structure, the memory cost is proportional to the amount of non-zero elements in
the representation vector. To further reduce such memory overhead, Jegou et al. proposed to approximate the orig- inal visual word occurrence vector by projecting it onto a set of pre-deï¬ned sparse projection functions, generat- ing multiple min-BOF descriptors [179]. Those min-BOF descriptors is further quantized for indexing. With similar attempt, in [16][180], min-Hash is proposed to describe images by mapping the visual word occurrence vector to a low-dimensional representation by a group of min-hash functions and deï¬ne image similarity as the visual word set overlap. Consequently, only a small constant amount of data per image need to be stored. The potential concern of min-hashing [16][180] and its variant [126] is that although high retrieval precision can be achieved, the retrieval recall performance may be limited unless many more hashing ta- bles are involved, which, however, imposes severe memory burden.
# 6 IMAGE SCORING
In multimedia retrieval, the target results in the index image database are assigned with a relevance score for ranking and then returned to users. The relevance score can be deï¬ned either by measuring distance between the aggregated fea- ture vectors of image representation or from the perspective of voting from relevant visual feature matches. | 1706.06064#63 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 65 | i=1 X where the feature aggregation vectors of image Iq and Im are denoted as [q1, q2, · · · , qN ] and [m1, m2, · · · , mN ], respectively, and N denotes the vector dimension. In [10], it is revealed that L1-norm yields better retrieval accuracy than L2-norm with the BoW model. Lin et al. extended the above feature distance to measure partial similarity between images with an optimization scheme [181].
When the BoW model is adopted for image representa- tion, the feature aggregation vector is essentially a weighted visual word histogram obtained based on the feature quan- tization results. To distinguish the signiï¬cance of visual words in different images, term frequency (TF) and inverted document/image frequency (IDF) are widely applied in many existing state-of-the-art algorithms [10][12][9][15][17]. Generally, the visual word vector weighted by TF and IDF are Lp-normalized for later distance computation. When the codebook size is much larger than the local feature amount in images, the aggregated feature vector of image is very sparse and we only need to check those visual words appearing in both images as illustrated in Eq. 6 [10], which is very efï¬cient in practical implementation.
10 | 1706.06064#65 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 66 | 10
N D(Iq, Im) = |qi â mi|p (5) i=1 X = 2 + (|qi â mi|p â qp i â mp i )(6) Xi|qi6=0,mi6=0
However, the dissimilarity measure by the Lp-distance is not optimal. As revealed in [182], there exists the neigh- borhood reversibility issue, which means that an image is usually not the k-nearest neighbor of its k-nearest neighbor images. Such issue causes that problem that some images are frequently returned while others are rarely returned when submitting query images. To address this problem, Jegou et al. proposed a novel contextual dissimilarity mea- sure to reï¬ne the Euclidean distance based distance [182]. It modiï¬es the neighborhood structure in the BoW space by iteratively estimating distance update terms in the spirit of Sinkhorns scaling algorithm. Alternatively, in [183], a probabilistic framework is proposed to model the feature to feature similarity measure and a query adaptive similarity is derived. Different from the above approaches, in [184], the similarity metric is implicitly learnt with diffusion processes by exploring the afï¬nity graphs to capture the intrinsic manifold of database images. | 1706.06064#66 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 67 | In [138], Jegou et al. investigated the phenomenon of co- missing and co-occurrence in the regular BoW vector repre- sentation. The co-missing phenomenon denotes a negative evidence, i.e., a visual word is jointly missing from two BoW vectors. To include the under-estimated evidence for similarity measurement reï¬nement, vectors of images are centered by mean substraction [138]. On the other hand, the co-occurrence of visual words across BoW vectors will cause over-counting of some visual patterns. To limit this impact, a whitening operation is introduced to the BoW vector to gen- erate a new representation [138]. Such preprocessing also applies to the VLAD vector [116]. Considerable accuracy gain has been demonstrated with the above operations.
# 6.2 Voting Based Scoring
In local feature based image retrieval, the image similarity is intrinsically determined by the feature matches between images. Therefore, it is natural to derive the image similarity score by aggregating votes from the matched features. In this way, the similarity score is not necessarily normalized, which is acceptable considering the nature of visual ranking in image retrieval. | 1706.06064#67 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 68 | In [13], the relevance score is simply deï¬ned by counting how many pairs of local feature are matches across two images. In [35], Jegou et al formulated the scoring function as a cumulation of squared TF-IDF weights on shared visual words, which is essentially a BOF (bag of features) inner product [35]. In [17], the image similarity is deï¬ned as the sum of the TF-IDF score [20], which is further enhanced with a weighting term by matching bundled feature sets. The weighting term consists of membership term and geometric term. The former term is deï¬ned as the number of shared visual words between two bundled features, while the latter is formulated using relative ordering to penalize geometric
inconsistency of the matching between two bundled fea- tures. In [185][186], Zheng et al propose a novel Lp-norm IDF to extend the classic IDF weighting scheme. | 1706.06064#68 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 69 | The context clues in the descriptor space and the spatial domain are important to contribute the similarity score when comparing images. In [123], a contextual weighting scheme is introduced to enhance the original IDF-based voting so as to improve the classic vocabulary tree approach. Two kinds of weighting scheme, i.e., descriptor contextual weighting (DCW) and spatial contextual weighting, are formulated to multiply the basic IDF weight as a new weighting scheme for image scoring. In [187], Shen et al. proposed a spatially-constrained similarity measure based on a certain transformation to formulate voting score. The transformation space is discretized and a voting map is gen- erated based on the relative locations of matched features to determine the optimal transformation.
In [179], each indexed feature is embedded with a binary signature and the image distance is deï¬ned as a summation of the hamming distance between matched features, of which the distance for the unobserved match is set as sta- tistical expectation of the distance. Similar scoring scheme for the unobserved match is also adopted by Liu et al. [157]. In [63], to tolerate the correspondences of multiple visual objects with different transformations, local similarity of de- formations is derived from the peak value in the histogram of pairwise geometric consistency [188]. This similarity score is used as a weighting term to the general voting scores from local correspondences. | 1706.06064#69 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 70 | In image retrieval with visual word representation, sim- ilar to text-based information retrieval [189], there is a phenomenon of visual word burstiness, i.e., some visual element appears much more frequently in an image than the statistically expectation, which undermines the visual similarity measure. To address this problem, Jegou et al proposed three strategies to penalize the voting scores from the bursting visual words by removing multiple local matches and weaken the inï¬uence of intra- and inter-images bursts [190] [191].
# 7 SEARCH RERANKING
re- The initially returned result ï¬ned by exploring the visual [193] or enhancing the original query. Geometric veriï¬ca- tion [11] [18] [12] [13] [126] [194], query expansion [14] [195], and retrieval fusion [24] are three of the most successful post-processing techniques to boost the accuracy of large scale image search. In the following, we will review the related literature in each category.
# 7.1 Geometric Context Veriï¬cation | 1706.06064#70 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 71 | # 7.1 Geometric Context Veriï¬cation
In image retrieval with local invariant features, the feature correspondences between query and database images are built based on the proximity of local features in the descrip- tor space. As a popular criteria, a tentative correspondence is built if the corresponding two local features are quantized to the same visual word of a pre-trained visual vocabulary. However, due to the ambiguity of local descriptor and the quantization loss, false correspondences of irrelevant visual
11
content are inevitably incurred, which confuse the similarity measurement for images and degrade the retrieval accuracy. Note that, besides the descriptor, local invariant features are characterised by other geometric context, such as the location of key points in image plane, orientation, scale, and spatial co-occurrences with other local features. Such geometric context is an important clue to depress or exclude those false matches.
Generally, among the inliers in the correspondences set, there is an underlying transformation model. If the model is uncovered, we can easily distinguish the inliers from the outliers. To model the transformation of visual object or scene across images, an afï¬ne transformation model with six parameters can be used, which estimates the rotation, scaling, translation, and perspective change in a single homography [11]. For some difï¬cult cases, there may exist multiple homographies which makes the model estimation problem much more challenging. | 1706.06064#71 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 72 | Some approaches estimate the transformation model in an explicit way to verify the correspon- dences. Those methods are either based the RANSAC-like idea [11][8][196] [63] or follow the Hough voting strat- egy [8][197]. The key idea of RANSAC [198] is to gen- erate hypotheses on random sets of correspondences and identify a geometric model with the maximum inliers. Sta- tistically speaking, the genuine model can be recovered with sufï¬cient number of correspondence sampling and model evaluation. However, when the rate of inliers is small, the expected number of correspondence sampling is large, which incurs high computational complexity. In [11], by adopting the region shape of local feature, a hypothe- sis is generated with single correspondence, which make it feasible to enumerate all hypotheses and signiï¬cantly reduces the computational cost compared with RANSAC. There are two issues on the RANSAC based algorithms. Firstly, it needs a parameter for hypothesis veriï¬cation, which is usually deï¬ned in an ad-hoc way. Secondly, the computational complexity is quadratic with | 1706.06064#72 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 73 | for hypothesis veriï¬cation, which is usually deï¬ned in an ad-hoc way. Secondly, the computational complexity is quadratic with respect to the number of correspondences, which is somewhat expensive. An alternative to the RANSAC-like methods, Hough voting strategy [8] [199] operates in a transformation space. In this case, the voting operation is linear to the correspon- dence number. In [12], the Hough voting is conducted in the space of scale and orientation. Based on the SIFT feature correspondences between images, it builds two histograms on the orientation difference and scale difference separately. Assuming that truly matched features will share similar orientation difference, it identiï¬es the peak points in the histogram on orientation difference of matched features and regard those feature pairs with orientation difference far from the peak as irrelevant and false matches. Simi- lar operation is also performed on the scale difference of matched features to further remove those geometrically inconsistent SIFT matches. In [20], Zhang et al. built a 2D Hough voting space based on the relative displacements of corresponding local features to derive the geometric- preserving visual phrase | 1706.06064#73 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 74 | Zhang et al. built a 2D Hough voting space based on the relative displacements of corresponding local features to derive the geometric- preserving visual phrase (GVP).This algorithm can be ex- tended to address the transformation invariance to scale and rotation with the price of high memory overhead to maintain the Hough histograms. The potential problem in Hough voting is the ï¬exibility issue in the deï¬nition of the | 1706.06064#74 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 75 | bin size for the transformation space partition. To address the problem, in [197], motivated by the pyramid matching scheme [200], Tolias et al. propose a Hough pyramid match- ing scheme. It approximates afï¬nity by bin size and group the correspondences based on the afï¬nity in a bottom-up way. Notably, the complexity of this algorithm is linear to the correspondence number. In [199], the Hough pyramid matching scheme is extended by including the soft assign- ment for feature quantization on the query image. Different from the above methods, Li et al. proposed a novel pair- wise geometric matching method [194] for implicit spatial veriï¬cation at a signiï¬cantly reduced computational cost. To reduce the correspondence redundancy, it ï¬rst builds the initial correspondence set with a one-versus-one matching strategy, which is further reï¬ned based on Hough voting in the scaling and rotation transformation space [12]. Based on the reliable correspondence set, a new pairwise weighting method is proposed to measure the matching score between two images. | 1706.06064#75 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 76 | Some other algorithms approach the geometric context veriï¬cation problem without explicit handling the transfor- mation model. Sivic et al. adopted the consistency of spatial context in local feature groups to verify correspondences [9]. In [18], a spatial coding scheme is proposed to encode into two binary maps by comparing the relative coordi- nates of matched feature points in horizontal and vertical directions, respectively. Then it recursively removes geo- metrically inconsistent matches by analyzing those maps. Although spatial coding map is invariant to image changes in translation and scaling, it cannot handle the rotation change. In [13] [201], Zhou et al. extended the spatial cod- ing by including the characteristic orientation and scale of SIFT feature and proposed two geometric context coding methods, i.e., geometric square coding and geometric fan coding. The geometric coding algorithm can well handle image changes in translation, rotation, and scaling. In [202], Chu et al. proposed a Combined-Orientation-Position (COP) consistency graph model to measure the relative spatial consistency among the candidate matches of SIFT features with a coarse-to-ï¬ne family of evenly sectored polar coor- dinate system. Those spatially inconsistent noisy features are effectively identiï¬ed and rejected by detecting the group of candidate feature matches with the largest average COP consistency.
# 7.2 Query Expansion | 1706.06064#76 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 77 | # 7.2 Query Expansion
Query expansion, leveraged from text retrieval, reissues the initially highly-ranked results to generate new queries. Some relevant features, which are not present in the original query, can be used to enrich the original query to further im- prove the recall performance. Several expansion strategies, such as average query expansion, transitive closure expan- sion, recursive average query expansion, intra-expansion, and inter-expansion, etc., have been discussed in [14] [195]. In [23], a discriminative query expansion algorithm is proposed. It takes spatially veriï¬ed images as positive data and images with low tf-idf scores as the negative training data. Then, a classiï¬er is learnt on-the-ï¬y and images are sorted by their signed distances from the decision boundary. In [203], Xie et al. constructed a sparse graph by connecting
12 | 1706.06064#77 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 78 | 12
potentially relevant images ofï¬ine and adopted a query- dependent algorithm, i.e., HITS [204], to reranking images based on afï¬nity propagation. Further, Xie et al. formu- lated the search process with a heterogeneous graph model and proposed two graph-based re-ranking algorithms to improve the search precision and recall, respectively [205]. It ï¬rst incrementally identiï¬es the most reliable images from the database to expand the query so as to boost the recall. After that, an image-feature voting scheme is used to iteratively update the scores of images and features to re- rank images. In [206], a contextual query expansion scheme is proposed to explore the common visual patterns. The contextual query expansion is performed in both the visual word level and the image level. | 1706.06064#78 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 79 | relevance feedback [1] has been demonstrated to be success- ful search re-ranking technique and well studied be- fore attention in recent years [207] [208] [209] [210] [211] [212]. In relevance feed- back, the key idea is to learn a query-speciï¬c similarity metric based on the relevant and irrelevant examples in- dicated by users. Some discriminative models are learned with SVM [207][208] or boosting schemes [213]. Considering that users are usually reluctant or impatient to specify positive or negative images, user click log information can be collected as feedback to implicitly improve the retrieval system [31] [214]. For more discussion on relevance feed- back, we refer readers to [215] [216] for a comprehensive survey.
# 7.3 Retrieval Fusion | 1706.06064#79 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 80 | # 7.3 Retrieval Fusion
An image can be represented by different features, based on which different methods can be designed for retrieval. If the retrieval results of different methods are comple- mentary to each other, they can be fused to obtain better results. Most approaches conduct retrieval fusion in the rank level. Fagin et al. proposed a rank aggregation algorithm to combine the image ranking lists of multiple independent retrieval methods or âvotersâ [217]. In [24], the retrieval fusion is formulated as a graph-based ranking problem. A weighted undirected graph is built based on the retrieval results of one method and the graphs corresponding to multiple retrieval methods are fused to a single graph, based on which, link analysis [218] or maximizing weighted density is conducted to identify the relevance score and rank the retrieval results. In [219], Ye et al. proposed a novel rank minimization method to fuse the conï¬dence scores of multiple different models. It ï¬rst constructs a comparative relationship matrix based on the predicted conï¬dent scores for each model. With the assumption that the relative score relations are consistent across different models with some sparse deviations, it formulates the score fusion problem as seeking a shred rank-2 matrix and derives a robust a score vector. | 1706.06064#80 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 81 | Different from the above fusion methods, Zheng et al. approached the retrieval fusion in the score level [103]. Motivated by the shape differences in the ranked score curve between good and bad representation features, it normalizes the score curves by reference curves trained on irrelevant data and derives an effectiveness score based
on the area under the normalized score curve. Then, the query similarity measurement is adaptively formulated in a product manner over the feature scores weighted by the effectiveness score.
# 8 DATASET AND PERFORMANCE EVALUATION
To quantitatively demonstrate the effectiveness and efï¬- ciency of various image retrieval algorithms, it is indispens- able to collect some benchmark datasets and deï¬ne the eval- uation metrics. In this section, we discuss the recent ground truth datasets and distractor datasets used in experimental study for image retrieval. Besides, we introduce the key evaluation indicators in CBIR, including accuracy, efï¬ciency, and memory cost.
# 8.1 Recent Dataset for CBIR | 1706.06064#81 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 82 | # 8.1 Recent Dataset for CBIR
Intuitively, the ground-truth dataset should be sufï¬cient large so as to well demonstrate the scalability of image retrieval algorithms. However, considering the tedious la- bor in dataset collection, the existing ground-truth dataset are relatively small, but mixed with random million-scale distractor database for evaluation on scalability. The exist- ing ground-truth datasets target on particular object/scene retrieval or partial-duplicate Web image retrieval. Generally, the ground-truth images contain a speciï¬c object or scene and may undergo various changes and be taken under different views or changes in illumination, scale, rotation, partial occlusion, compression rate, etc. Typical ground truth dataset for this task includes the UKBench dataset [10], the Oxford Building dataset [11], and the Holidays dataset [12], etc. MIR Flickr-1M and Flickr-1M are two different million- scale databases which are usually used as distractor to evaluate the scalability of image retrieval algorithms. For convenience of comparison and reference, we list the gen- eral information of those recent datasets popularly used in CBIR in Table 1. Some sample images from those datasets are shown in Fig. 3. | 1706.06064#82 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 83 | UKBench dataset It contains 10,200 images from 2,550 categories9. In each category, there are four images taken on the same scene or object from different views or illumination conditions. All the 10,200 images are taken as query and their retrieval performances are averaged.
Holidays dataset There are 1,491 images from 500 groups in the Holidays dataset10. Images in each group are taken on a scene or an object with various viewpoints. The ï¬rst image in each group is selected as query for evaluation. Oxford Building dataset (Oxford-5K) The Oxford Build- ings Dataset11 consists of 5062 images collected from Flickr12 by searching for particular Oxford landmarks. The collection has been manually annotated to generate a comprehensive ground truth for 11 different landmarks, each represented by 5 possible queries. This gives a set of 55 queries over which an object retrieval system can be evaluated. Some junk images are mixed in it as distractor.
9. http://www.vis.uky.edu/â¼stewe/ukbench/ 10. http://lear.inrialpes.fr/people/jegou/data.php 11. http://www.robots.ox.ac.uk/â¼vgg/data/oxbuildings/ 12. http://www.ï¬ickr.com/
13 | 1706.06064#83 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 84 | 13
TABLE 1 General information of the popular retrieval datasets in CBIR. The âmixedâ database type denotes that the corresponding dataset is a ground truth dataset mixed with distractor images.
Resolution 640 Ã 480 1024 Ã 768 1024 Ã 768 1024 Ã 768 460 Ã 350 (average) 1024 Ã 768 1000 Ã 720 (average) 320 Ã 240 640 Ã 480 500 Ã 500 N/A Category Number 2,550 500 11 12 33 32 200 200 1,200 N/A N/A Database Name Database Type Database Size Query Number UKBench Holidays Oxford-5K Paris DupImage FlickrLogos-32 INSTRE ZuBuD SMVS MIR Flickr-1M Flickr1M 10,200 1,491 6,053 6,412 1,104 8,240 28,543 1,005 1,200 1,000,000 1,000,000 10,200 500 55 500 108 500 N/A 115 3,300 N/A N/A Ground Truth Ground Truth Mixed Mixed Ground Truth Mixed Ground Truth Ground Truth Ground Truth Distractor Distractor
Paris dataset In the Paris dataset13, there are 6,412 im- ages, which are collected from Flickr by searching for 12 text queries of particular Paris landmarks. For this dataset, 500 query images are used for evaluation. | 1706.06064#84 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 85 | DupImage dataset This dataset contains 1,104 images from 33 groups14. Each group corresponds to a logo, a painting, or an artwork, such as KFC, American Gothic Painting, Mona Lisa, etc. 108 representative query images are selected from those groups for evaluation.
MIR Flickr-1M This is a distractor dataset19, with one million images randomly downloaded from Flickr and re- sized to be no larger than 500 by 500.
Flickr1M is another distractor database containing SIFT features20 of one million images arbitrarily retrieved from Flickr. The original images in this database are not available.
# 8.2 Performance Evaluation for CBIR
FlickrLogos-32 dataset This dataset15 contains logo im- ages of 32 different brands which are downloaded from Flickr. All logo images in this dataset have an approximately planar structure. The dataset is partitioned into three subsets for evaluation, i.e., training set, validation set, and query set [220]. Of those 8,240 images in the dataset, 6,000 images contain no logos and works as distractors.
INSTRE As an instance-level benchmark dataset, the INSTRE dataset 16 contains two subsets, i.e., INSTRE-S and INSTRE-M [221]. In the former subset, there are 23,070 images, each with a single label of 200 classes. The latter subset contains 5,473 images and each image contains two instances from 100 object categories. | 1706.06064#85 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 86 | ZuBuD dataset The basic dataset contains 1,005 images of 201 buildings in Zurich, with 5 views for each building17. Besides, there are additional 115 query images which are not included in the basic dataset. The resolution of those images are uniformly 320 Ã 240.
Stanford Mobile Visual Search (SMVS) Dataset This dataset18 is targeted for mobile visual search and contains images taken by camera phone on products, CDs, books, outdoor landmarks, business cards, text documents, mu- seum paintings and video clips. It is characterized by rigid objects, widely varying lighting conditions, perspective dis- tortion, foreground and background clutter, and realistic ground-truth reference data [222]. In the dataset, there are 1,200 distinct categories. For each category, one reference image with resolution quality is collected for evaluation. There are 3,300 query images in total which are collected from heterogeneous low and high-end camera phones. | 1706.06064#86 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 87 | 13. http://www.robots.ox.ac.uk/â¼vgg/data/parisbuildings/ 14. http://pan.baidu.com/s/1jGETFUm 15. http://www.multimedia-computing.de/ï¬ickrlogos/ 16. http://vipl.ict.ac.cn/isia/instre/ 17. http://www.vision.ee.ethz.ch/showroom/zubud/index.en.html 18. http://purl.stanford.edu/rb470rw0983
In the design of a multimedia content-based retrieval sys- tem, there are three key indicators which should be carefully considered: accuracy, efï¬ciency, and memory cost. Usually, a retrieval method contributes to improving at least one of those indicators with little sacriï¬ce in the other indicators. | 1706.06064#87 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 88 | Accuracy To measure the retrieval quality quantitatively, the database images are categorized into difference rele- vance levels and the accuracy score is summarized based on the rank order of the database images. For different relevance levels, there are different accuracy metrics. Where there are only two relevance level, i.e., relevant and irrel- evant, average precision (AP) is widely used to evaluate the retrieval quality of a single queryâs retrieval results. AP takes consideration of both precision and recall. Precision denotes the fraction of retrieved (top k) images that are relevant while recall means fraction of relevant image that are retrieved (in the top k returned results). Generally, for a retrieval system, precision decreases as either the number of images retrieved increases or recall grows. AP averages the precision values from the rank positions where a relevant image was retrieved, as deï¬ned in Eq. 7. To summarize the retrieval quality over multiple query images, the mean average precision (mAP) is usually adopted, which average the average precision over all queries.
AP = n k=1 P (k) · rel(k) R (7) | 1706.06064#88 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 89 | # P
where R denotes the number of relevant results for the current query image, P (k) denotes the precision of top k retrieval results, rel(k) is a binary indicator function equalling 1 when the k-th retrieved result is relevant to the current query image and 0 otherwise, and n denotes the total number of retrieved results.
19. http://medialab.liacs.nl/mirï¬ickr/mirï¬ickr1m/ 20. http://bigimbaz.inrialpes.fr/herve/siftgeo1M/
14
Fig. 3. Samples images of the existing datasets. First row: UKBench dataset; second row: Holidays dataset; third row: Oxford Building dataset; fourth row: DupImage dataset; ï¬fth row: INSTRE dataset; sixth row: ZuBuD dataset; seventh row: SMVS dataset.
When there are multiple relevance levels, we can resort to normalized discounted cumulative gain (NDCG) metric deï¬ned in Eq. 8 to summarize the ranking results.
N DCG = 1 N (r1 + n f (rk) log2(k) ), (8) | 1706.06064#89 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 90 | # k=2 X
where n denotes the number of retrieved images, rk denotes the relevance level, f (·) is function to tune the contribution of difference relevance levels, and N denotes the normalized term to ensure that the NDCG score for the ideal retrieved results is 100%. Popular deï¬nitions of f (·) include f (x) = x
and f (x) = 2x â1, with the latter to emphasize on retrieving highly relevant images.
Besides the above measures, some simple measures may be adopted for special dataset. In the public UKBench dataset, considering that there are four relevant images for all queries, the N-S score, i.e., the average 4 times top-4 precision over the dataset, are used to measure the retrieval accuracy [10].
Computational Efï¬ciency The efï¬ciency of a retrieval system involves the time cost in visual vocabulary (code- book) construction, visual feature indexing, and image querying. The ï¬rst two items are performed off-line, while
15
the last one is conducted on-line. Both the off-line and on- line processing is expected to be as fast as possible. Specially, the on-line querying is usually expected to be responded in real time. | 1706.06064#90 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 91 | In a multimedia content-based visual retrieval system, the memory cost usually refers to the memory usage in the on-line query stage. Generally, the memory is mainly spent on the quantizer and the index ï¬le of database, which need to be loaded into the main memory for on-line retrieval. Popular quantizer includes tree-based structure, such as hierarchical vocabulary tree, randomized forests, etc, which usually cost a few hundred mega-bytes memory for codebook containing million-scale visual words. In some binary code based quantization meth- ods [36] [72], the quantizer is simple hash function with negligible memory overhead. For the index ï¬le, the memory cost is proportional to the indexed database size. When the database images are represented by local features and each local feature is indexed locally, the index ï¬le is proportional to the amount of indexed features and the memory cost per indexed feature.
9 FUTURE DIRECTIONS Despite the extensive research efforts in the past decade, there is still sufï¬cient space to further boost content based visual search. In the following, we will discuss several directions for future research, on which new advance shall be made in the next decade.
# 9.1 Ground-Truth Dataset Collection | 1706.06064#91 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 92 | # 9.1 Ground-Truth Dataset Collection
In the multimedia and computer vision ï¬eld, ground-truth datasets are motivated by the speciï¬c tasks. At the begin- ning of those dataset construction, they inspire researchers to update the performance records with their best efforts, leading to many classic ideas and algorithms to address the research problem. However, with the advance to address those datasets, the break-through of some algorithms may suffer from the over-ï¬tting to the dataset. Meanwhile, with deeper understanding and clearer deï¬nition of the research problem, the limitation of existing datasets is revealed and new datasets are expected. For content-based image re- trieval, we also expect better ground-truth dataset to be collected and released. Generally, the new ground-truth datasets shall be speciï¬c to eliminate the ambiguity of rele- vance of image content, such as logo datasets. Meanwhile, the scale of the dataset shall be sufï¬ciently large so as to distinguish the problem of CBIR from image classiï¬cation.
# 9.2 Intention Oriented Query Formation and Selection | 1706.06064#92 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 93 | # 9.2 Intention Oriented Query Formation and Selection
Intention gap is the ï¬rst and of the greatest challenge in content-based image retrieval. A simple query in the form of example, color map or sketch map is still insufï¬cient in most time to reï¬ect the user intention, consequently generating unsatisfactory retrieval results. Besides the traditional query formations, assistance from user to specify the concrete ex- pectation will greatly alleviate the difï¬culty of the following image retrieval process. Considering that the end-users may be reluctant to involve much in the query formation, it is still possible to design convenient query formation interface
to reduce the user involvement as much as possible. For instance, it is easy for a user to specify the region of interest in an example image for retrieval, or indicate the expected results are partial-duplicates or just similar in spatial color and texture. It is also possible to predict the potential inten- tions based on the initial query and make conï¬rmation with end-user. In all, rather than passively induce the intension behind the query, it is beneï¬cial to actively involve end-user in the retrieval process. | 1706.06064#93 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 94 | In image retrieval, the search performance is signiï¬- cantly impacted by the quality of the query. How to select a suitable query towards the optimal retrieval is a nontrivial issue. The query quality is related with many factors, includ- ing resolution, noise pollution, afï¬ne distortion, background clutter, etc. In the scenario of mobile search, the query can be selected by guiding the end user to retake better photos. In the server end, automatic retrieval quality assessment methods [223] [224] can be designed to select potential candidate from the initial retrieval results of high precision.
# 9.3 Deep Learning in CBIR
Despite the advance in content-based visual retrieval, there is still signiï¬cant gap towards semantic-aware retrieval from visual content. This is essentially due to the fact that current image representation schemes are hand-crafted and insuf- ï¬cient to capture the semantics. Due to the tremendous diversity and quantity in multimedia visual data, most existing methods are un-supervised. To proceed towards semantic-aware retrieval, scalable scalable supervised or semi-supervised learning are promising to learn semantic- aware representation so as to boost the content-based re- trieval quality. The success of deep learning in large-scale visual recognition [99] [96] [95] [225] has already demon- strated such potential. | 1706.06064#94 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 95 | To adapt those existing deep learning techniques to CBIR, there are several non-trivial issues that deserve re- search efforts. Firstly, the learned image representation with deep learning shall be ï¬exible and robust to various com- mon changes and transformations, such as rotation and scal- ing. Since the existing deep learning relies on the convolu- tional operation with anisotropic ï¬lters to convolve images, the resulted feature maps are sensitive to large translation, rotation, and scaling changes. It is still an open problem as whether that can solved by simply including more training samples with diverse transformations. Secondly, since com- putational efï¬ciency and memory overhead are emphasized in particular in CBIR, it would be beneï¬cial to consider those constraints in the structure design of deep learn- ing networks. For instance, both compact binary semantic hashing codes [59] [65] and very sparse semantic vector representations are desired to represent images, since the latter are efï¬cient in both distance computing and memory storing while the latter is well adapted to the inverted index structure.
# 9.4 Unsupervised Database Mining
In traditional content-based image retrieval algorithms and systems, the database images are processed independently without considering their potential relevance context in- formation. This is primarily due to the fact that, there is
16 | 1706.06064#95 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 96 | In traditional content-based image retrieval algorithms and systems, the database images are processed independently without considering their potential relevance context in- formation. This is primarily due to the fact that, there is
16
usually no label information for the database images and the potential category number is unlimited. Those constraints limit the application of sophisticated supervised learning algorithms in CBIR. However, as long as the database is large, it is likely that there exist some subsets of images and images in each sub-set are relevant to each other images. Therefore, it is feasible to explore the database images with some unsupervised techniques to uncover those sub-sets in the off-line processing stage. If we regard each database image as a node and the relevance level between images as edge to link images, the whole image database can be repre- sented as a large graph. Then, the sub-sets mining problem can be formulated as a sub-graph discovery problem. On the other hand, in practice, new images may be incrementally included into the graph, which casts challenge to dynami- cally uncover those sub-graphs on the ï¬y. The mining results in the off-line stage will be beneï¬cial for the on-line query to yield better retrieval results.
# 9.5 Cross-modal Retrieval | 1706.06064#96 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 97 | # 9.5 Cross-modal Retrieval
In the above discussion of this survey, we focus on the visual content for image retrieval. However, besides the visual features, there are other very useful clues, such as the textual information around images in Web pages, the click log of users when using the search engines, the speech information in videos, etc. Those multi-modal clues are complementary to each to collaboratively identify the visual content of images and videos. Therefore, it would be beneï¬cial to explore cross-modal retrieval and fuse those multi-modal features with different models. With multi- modal representation, there are still many open search topics in terms of collaborative quantization, indexing, search re- ranking, etc.
# 9.6 End-to-End Retrieval Framework | 1706.06064#97 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 98 | As discussed in the above sections, the retrieval framework is involved with multiple modules, including feature ex- traction, codebook learning, feature quantization, feature quantization, image indexing, etc. Those modules are in- dividually designed and independently optimized for the retrieval task. On the other hand, if we investigate the structure of the convolutional neural network (CNN) in deep learning, we can ï¬nd a very close analogy between the BoW model and the CNN model. The convolutional ï¬lters used in the CNN model works in a similar way as the code- words of the codebook in the BoW model. The convolution results between the image patch and the convolution ï¬lter are essentially the soft quantization results, with the max- pooling operation similar to the local aggregation in the BoW model. As long as the learned feature vector is sparse, we can also adopt the inverted index structure to efï¬ciently index the image database. Different from the BoW model, the above modules in the CNN model are collaboratively optimized for the task of image classiï¬cation. Based on the above analogy, similarly, we may also resort to an end-to- end paradigm to design a framework that takes images as input and outputs the index-oriented features directly, with the traditional key retrieval-related modules implicitly and collaboratively optimized.
17 | 1706.06064#98 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 99 | 17
# 9.7 Social Media Mining with CBIR
Different from the traditional unstructured Web media, the emerging social media in recent years have been charac- terized by community based personalized content creation, sharing, and interaction. There are many successful promi- nent platforms of social media, such as Facebook, Twitter, Wikipedia, LinkedIn, Pinterest, etc. The social media is enriched with tremendous information which dynamically reï¬ects the social and cultural background and trend of the community. Besides, it also reveals the personal affection and behavior characteristics. As an important media of the user-created content, the visual data can be used as an entry point with the content-based image retrieval technique to uncover and understand the underlying community struc- ture. It would be beneï¬cial to understand the behavior of individual users and conduct recommendation of products and services to users. Moreover, it is feasible to analyze the sentiment of crowd for supervision and forewarning.
# 9.8 Open Grand Challenge | 1706.06064#99 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 100 | # 9.8 Open Grand Challenge
Due to the difference in deployment structure and availabil- ity of data, the research on content based image retrieval in the academia suffers a gap from the real application in industry. To bridge this gap, it is beneï¬cial to initiate some open grand challenges from the industry and involve the researchers in the academia to investigate the key difï¬culties in real scenarios. In the past ï¬ve years, there are some limited open grand challenge, such as the Microsoft Image Grand Challenge on Image Retrieval21 and Alibaba Large- Scale Image Search Challenge22. In the future, we would expect many more such grand challenges. The open grand challenge will only only advance the research progress in the academia, but also beneï¬t the industry with more and better practical and feasible solutions to the real-world challenges.
# 10 CONCLUSIONS
In this paper, we have investigated the advance on content- based image retrieval in recent years. We focus on the ï¬ve key modules of the general framework, i.e., query formation, image representation, image indexing, retrieval scoring, and search re-ranking. For each component, we have discussed the key problems and categorized a variety of representative strategies and methods. Further, we have summarized eight potential directions that may boost the advance of content based image retrieval in the near future.
# REFERENCES
[1]
[2] | 1706.06064#100 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 101 | # REFERENCES
[1]
[2]
Y. Rui, T. S. Huang, M. Ortega, and S. Mehrotra, âRelevance feed- back: a power tool for interactive content-based image retrieval,â IEEE Transactions on Circuits and Systems for Video Technology, vol. 8, no. 5, pp. 644â655, 1998. A. Alzubi, A. Amira, and N. Ramzan, âSemantic content-based image retrieval: A comprehensive study,â Journal of Visual Com- munication and Image Representation, vol. 32, pp. 20â54, 2015.
21. http://acmmm13.org/submissions/call-for-multimedia-grand- challenge-solutions/msr-bing-grand-challenge-on-image-retrieval- scientiï¬c-track
22. http://tianchi.aliyun.com/competition/introduction.htm?spm=5176.100069 .5678.1.SmufkG&raceId=231510& lang=en US | 1706.06064#101 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 102 | X. Li, T. Uricchio, L. Ballan, M. Bertini, C. G. Snoek, and A. D. Bimbo, âSocializing the semantic gap: A comparative survey on image tag assignment, reï¬nement, and retrieval,â ACM Comput- ing Surveys (CSUR), vol. 49, no. 1, p. 14, 2016. Z. Lin, G. Ding, M. Hu, and J. Wang, âSemantics-preserving hashing for cross-view retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3864â3872. A. W. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain, âContent-based image retrieval at the end of the early years,â IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, no. 12, pp. 1349â1380, 2000. | 1706.06064#102 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 103 | [6] M. S. Lew, N. Sebe, C. Djeraba, and R. Jain, âContent-based mul- timedia information retrieval: State of the art and challenges,â ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), vol. 2, no. 1, pp. 1â19, 2006. Y. Liu, D. Zhang, G. Lu, and W.-Y. Ma, âA survey of content- based image retrieval with high-level semantics,â Pattern Recog- nition, vol. 40, no. 1, pp. 262â282, 2007. D. G. Lowe, âDistinctive image features from scale invariant keypoints,â International Journal of Computer Vision, vol. 60, no. 2, pp. 91â110, 2004. J. Sivic and A. Zisserman, âVideo Google: A text retrieval ap- proach to object matching in videos,â in IEEE Conference on Computer Vision and Pattern Recognition, 2003, pp. 1470â1477. [10] D. Nister and H. Stewenius, âScalable recognition with a vocab- ulary tree,â in IEEE Conference on Computer Vision and Pattern | 1706.06064#103 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 105 | [8]
[12] H. Jegou, M. Douze, and C. Schmid, âHamming embedding and weak geometric consistency for large scale image search,â in European Conference on Computer Vision, 2008, pp. 304â317. [13] W. Zhou, H. Li, Y. Lu, and Q. Tian, âLarge scale image search with geometric coding,â in ACM International Conference on Multimedia, 2011, pp. 1349â1352.
[14] O. Chum, J. Philbin, J. Sivic, M. Isard, and A. Zisserman, âTotal recall: Automatic query expansion with a generative feature model for object retrieval,â in International Conference on Computer Vision, 2007, pp. 1â8. J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman, âLost in quantization: Improving particular object retrieval in large scale image databases,â in IEEE Conference on Computer Vision and Pattern Recognition, 2008, pp. 1â8.
[16] O. Chum, J. Philbin, and A. Zisserman, âNear duplicate image detection: min-hash and tf-idf weighting,â in British Machine Vision Conference, vol. 3, 2008, p. 4. | 1706.06064#105 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 106 | [17] Z. Wu, Q. Ke, M. Isard, and J. Sun, âBundling features for large scale partial-duplicate web image search,â in IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 25â32.
[18] W. Zhou, Y. Lu, H. Li, Y. Song, and Q. Tian, âSpatial coding for large scale partial-duplicate web image search,â in ACM International Conference on Multimedia, 2010, pp. 511â520.
[19] O. Chum, A. Mikulik, M. Perdoch, and J. Matas, âTotal recall II: Query expansion revisited,â in IEEE Conference on Computer Vision and Pattern Recognition, 2011, pp. 889â896.
[20] Y. Zhang, Z. Jia, and T. Chen, âImage retrieval with geometry- preserving visual phrases,â in IEEE Conference on Computer Vision and Pattern Recognition, 2011, pp. 809â816. | 1706.06064#106 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 107 | [21] X. Zhang, L. Zhang, and H.-Y. Shum, âQsrank: Query-sensitive hash code ranking for efï¬cient Ç«-neighbor search,â in IEEE Con- ference on Computer Vision and Pattern Recognition, 2012, pp. 2058â 2065. J. He, J. Feng, X. Liu, T. Cheng, T.-H. Lin, H. Chung, and S.- F. Chang, âMobile product search with bag of hash bits and boundary reranking,â in IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 3005â3012.
[23] R. Arandjelovic and A. Zisserman, âThree things everyone should know to improve object retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 2911â2918. S. Zhang, M. Yang, T. Cour, K. Yu, and D. N. Metaxas, âQuery speciï¬c fusion for image retrieval,â in European Conference on Computer Vision (ECCV), 2012.
[24]
[25] Q. Tian, S. Zhang, W. Zhou, R. Ji, B. Ni, and N. Sebe, âBuilding descriptive and discriminative visual codebook for large-scale | 1706.06064#107 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 108 | image applications,â Multimedia Tools and Applications, vol. 51, no. 2, pp. 441â477, 2011.
[26] W. Zhou, H. Li, Y. Lu, and Q. Tian, âLarge scale partial-duplicate image retrieval with bi-space quantization and geometric consis- tency,â in IEEE International Conference Acoustics Speech and Signal Processing, 2010, pp. 2394â2397. S. Zhang, Q. Tian, G. Hua, Q. Huang, and S. Li, âDescriptive visual words and visual phrases for image applications,â in ACM International Conference on Multimedia, 2009, pp. 75â84. S. Zhang, Q. Huang, G. Hua, S. Jiang, W. Gao, and Q. Tian, âBuilding contextual visual vocabulary for large-scale image ap- plications,â in ACM International Conference on Multimedia, 2010, pp. 501â510.
[27]
[28]
[29] W. Zhou, Q. Tian, Y. Lu, L. Yang, and H. Li, âLatent visual context learning for web image applications,â Pattern Recognition, vol. 44, no. 10, pp. 2263â2273, 2011. | 1706.06064#108 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 109 | [30] G. Tolias, Y. Avrithis, and H. Jgou, âTo aggregate or not to aggre- gate: selective match kernels for image search,â in International Conference on Computer Vision (ICCV), 2013.
[31] L. Zhang and Y. Rui, âImage searchÅfrom thousands to billions in 20 years,â ACM Transactions on Multimedia Computing, Communi- cations, and Applications (TOMM), vol. 9, no. 1s, p. 36, 2013. [32] X. Tang, K. Liu, J. Cui, F. Wen, and X. Wang, âIntentsearch: Cap- turing user intention for one-click internet image search,â IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 34, no. 7, pp. 1342â1353, 2012.
[33] B. Moghaddam, Q. Tian, N. Lesh, C. Shen, and T. S. Huang, âVisualization and user-modeling for browsing personal photo libraries,â International Journal of Computer Vision (IJCV), vol. 56, no. 1-2, pp. 109â130, 2004. | 1706.06064#109 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 110 | [34] R. Datta, D. Joshi, J. Li, and J. Z. Wang, âImage retrieval: Ideas, inï¬uences, and trends of the new age,â ACM Computing Surveys (CSUR), vol. 40, no. 2, p. 5, 2008.
[35] H. J´egou, M. Douze, and C. Schmid, âImproving bag-of-features for large scale image search,â International Journal of Computer Vision, vol. 87, no. 3, pp. 316â336, 2010.
[36] W. Zhou, Y. Lu, H. Li, and Q. Tian, âScalar quantization for large scale image search,â in ACM International Conference on Multimedia, 2012, pp. 169â178.
[37] Y. Cao, H. Wang, C. Wang, Z. Li, L. Zhang, and L. Zhang, âMindï¬nder: interactive sketch-based image search on millions of images,â in ACM International Conference on Multimedia (MM), 2010, pp. 1605â1608. | 1706.06064#110 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 111 | [38] C. Xiao, C. Wang, L. Zhang, and L. Zhang, âSketch-based image retrieval via shape words,â in ACM International Conference on Multimedia Retrieval (ICMR). ACM, 2015, pp. 571â574.
[39] P. Sousa and M. J. Fonseca, âSketch-based retrieval of drawings using spatial proximity,â Journal of Visual Languages & Computing, vol. 21, no. 2, pp. 69â80, 2010.
[40] M. J. Fonseca, A. Ferreira, and J. A. Jorge, âSketch-based retrieval of complex drawings using hierarchical topology and geometry,â Computer-Aided Design, vol. 41, no. 12, pp. 1067â1081, 2009. S. Liang and Z. Sun, âSketch retrieval and relevance feedback with biased svm classiï¬cation,â Pattern Recognition Letters, vol. 29, no. 12, pp. 1733â1741, 2008.
[41] | 1706.06064#111 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 112 | [41]
[42] Y. Cao, C. Wang, L. Zhang, and L. Zhang, âEdgel index for large- scale sketch-based image search,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011, pp. 761â768. J. Wang and X.-S. Hua, âInteractive image search by color map,â ACM Transactions on Intelligent Systems and Technology (TIST), vol. 3, no. 1, p. 12, 2011.
[44] H. Xu, J. Wang, X.-S. Hua, and S. Li, âImage search by concept map,â in International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2010, pp. 275â282.
[45] ââ, âInteractive image search by 2d semantic map,â in Interna- tional Conference on World Wide Web (WWW). ACM, 2010, pp. 1321â1324.
[46] T. Lan, W. Yang, Y. Wang, and G. Mori, âImage retrieval with structured object queries using latent ranking svm,â in European Conference on Computer Vision (ECCV). Springer, 2012, pp. 129â 142. | 1706.06064#112 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 113 | [47] G. Kim, S. Moon, and L. Sigal, âRanking and retrieval of image sequences from multiple paragraph queries,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1993â 2001.
18
[48] C. Wengert, M. Douze, and H. J´egou, âBag-of-colors for improved image search,â in ACM International Conference on Multimedia. ACM, 2011, pp. 1437â1440. J. Xie, Y. Fang, F. Zhu, and E. Wong, âDeepshape: Deep learned shape descriptor for 3d shape matching and retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1275â1283.
[49] | 1706.06064#113 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 114 | [49]
[50] F. Wang, L. Kang, and Y. Li, âSketch-based 3d shape retrieval using convolutional neural networks,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1875â 1883. S. Bai, X. Bai, Z. Zhou, Z. Zhang, and L. Jan Latecki, âGift: A real- time and scalable 3d shape search engine,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 5023â 5032.
[52] M. Park, J. S. Jin, and L. S. Wilson, âFast content-based image retrieval using quasi-gabor ï¬lter and reduction of image feature dimension,â in IEEE Southwest Symposium on Image Analysis and Interpretation.
[53] X.-Y. Wang, B.-B. Zhang, and H.-Y. Yang, âContent-based image retrieval by integrating color and texture features,â Multimedia Tools and Applications (MTA), vol. 68, no. 3, pp. 545â569, 2014. | 1706.06064#114 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 115 | [54] B. Wang, Z. Li, M. Li, and W.-Y. Ma, âLarge-scale duplicate detection for web image search,â in IEEE International Conference on Multimedia and Expo (ICME).
[55] C. Siagian and L. Itti, âRapid biologically-inspired scene clas- siï¬cation using features shared with visual attention,â IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 29, no. 2, pp. 300â312, 2007.
[56] B. Kulis and K. Grauman, âKernelized locality-sensitive hashing for scalable image search,â in International Conference on Computer Vision, 2009, pp. 2130â2137.
[57] Y. Weiss, A. Torralba, and R. Fergus, âSpectral hashing,â in Advances in Neural Information Processing Systems (NIPS), 2009, pp. 1753â1760.
[58] H. J´egou, M. Douze, and C. Schmid, âProduct quantization for nearest neighbor search,â IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 1, pp. 117â128, 2011. | 1706.06064#115 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 116 | [59] A. Torralba, R. Fergus, and Y. Weiss, âSmall codes and large image databases for recognition,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[60] D. G. Lowe, âObject recognition from local scale-invariant fea- tures,â in IEEE International Conference on Computer Vision, vol. 2, 1999, pp. 1150â1157. J. Matas, O. Chum, M. Urban, and T. Pajdla, âRobust wide- baseline stereo from maximally stable extremal regions,â Image and Vision Computing, vol. 22, no. 10, pp. 761â767, 2004.
[62] K. Mikolajczyk and C. Schmid, âScale & afï¬ne invariant interest point detectors,â International Journal of Computer Vision, vol. 60, no. 1, pp. 63â86, 2004. | 1706.06064#116 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 117 | [63] H. Xie, K. Gao, Y. Zhang, S. Tang, J. Li, and Y. Liu, âEfï¬cient feature detection and effective post-veriï¬cation for large scale near-duplicate image search,â IEEE Transactions on Multimedia (TMM), vol. 13, no. 6, pp. 1319â1332, 2011.
[64] E. Rosten, R. Porter, and T. Drummond, âFaster and better: A machine learning approach to corner detection,â IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 32, no. 1, pp. 105â119, 2010.
[65] A. Krizhevsky and G. E. Hinton, âUsing very deep autoencoders for content-based image retrieval,â in ESANN. Citeseer, 2011.
[66] Z. Wu, Q. Ke, J. Sun, and H.-Y. Shum, âA multi-sample, multi- tree approach to bag-of-words image representation for image retrieval,â in IEEE International Conference on Computer Vision. IEEE, 2009, pp. 1992â1999. | 1706.06064#117 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 118 | [67] H. Bay, T. Tuytelaars, and L. Van Gool, âSURF: Speeded up robust features,â in European Conference on Computer Vision, 2006, pp. 404â417.
[68] L. Zheng, S. Wang, Z. Liu, and Q. Tian, âPacking and padding: Coupled multi-index for accurate image retrieval,â in IEEE Con- ference on Computer Vision and Pattern Recognition, 2014.
[69] W. Zhou, H. Li, R. Hong, Y. Lu, and Q. Tian, âBSIFT: towards data-independent codebook for large scale image search,â IEEE Transactions on Image Processing (TIP), vol. 24, no. 3, pp. 967â979, 2015.
[70] Z. Liu, H. Li, L. Zhang, W. Zhou, and Q. Tian, âCross-indexing of binary SIFT codes for large-scale image search,â IEEE Transactions on Image Processing (TIP), 2014. | 1706.06064#118 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 119 | [71] G. Yu and J.-M. Morel, âAsift: an algorithm for fully afï¬ne invariant comparison,â Image Processing On Line, vol. 2011, 2011. [72] W. Dong, Z. Wang, M. Charikar, and K. Li, âHigh-conï¬dence near-duplicate image detection,â in ACM International Conference on Multimedia Retrieval (ICMR). ACM, 2012, p. 1.
[73] M. Calonder, V. Lepetit, C. Strecha, and P. Fua, âBrief: binary robust independent elementary features,â in European Conference on Computer Vision (ECCV), 2010, pp. 778â792.
[74] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, âOrb: an efï¬cient alternative to sift or surf,â in International Conference on Computer Vision, 2011, pp. 2564â2571. | 1706.06064#119 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 120 | [75] A. Alahi, R. Ortiz, and P. Vandergheynst, âFreak: fast retina keypoint,â in IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 510â517. S. Leutenegger, M. Chli, and R. Y. Siegwart, âBrisk: binary ro- bust invariant scalable keypoints,â in International Conference on Computer Vision, 2011, pp. 2548â2555. S. Zhang, Q. Tian, Q. Huang, W. Gao, and Y. Rui, âUSB: Ultra- short binary descriptor for fast visual matching and retrieval,â IEEE Transactions on Image Processing (TIP), vol. 23, no. 8, pp. 3671â3683, 2014. S. Madeo and M. Bober, âFast, compact and discriminative: Evaluation of binary descriptors for mobile applications,â IEEE Transactions on Multimedia, 2016. S. Zhang, Q. Tian, K. Lu, Q. Huang, and W. Gao, âEdge-SIFT: Discriminative binary descriptor for scalable partial-duplicate mobile search,â | 1706.06064#120 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 122 | [77]
[79]
[81] M. Douze, A. Ramisa, and C. Schmid, âCombining attributes and ï¬sher vectors for efï¬cient image retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2011, pp. 745â752. S. Zhao, H. Yao, Y. Yang, and Y. Zhang, âAffective image retrieval via multi-graph learning,â in ACM International Conference on Multimedia (MM). ACM, 2014, pp. 1025â1028.
[83] R. Tao, A. W. Smeulders, and S.-F. Chang, âAttributes and cate- gories for generic instance search from one example,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 177â186.
[84] A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth, âDescribing objects by their attributes,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR). | 1706.06064#122 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 124 | [86] L. Torresani, M. Szummer, and A. Fitzgibbon, âEfï¬cient object category recognition using classemes,â in European Conference on Computer Vision (ECCV). Springer, 2010, pp. 776â789. J. Deng, A. C. Berg, and L. Fei-Fei, âHierarchical semantic in- dexing for large scale image retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2011, pp. 785â792. J. Cai, Z.-J. Zha, M. Wang, S. Zhang, and Q. Tian, âAn attribute- assisted reranking model for web image search,â IEEE Transac- tions on Image Processing (TIP), vol. 24, no. 1, pp. 261â272, 2015. S. Zhang, M. Yang, X. Wang, Y. Lin, and Q. Tian, âSemantic-aware co-indexing for image retrieval,â in IEEE International Conference on Computer Vis, 2013. S. Karayev, M. Trentacoste, H. Han, A. Agarwala, T. Darrell, A. | 1706.06064#124 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 126 | [87]
[89]
[90]
[91] T. Hofmann, âUnsupervised learning by probabilistic latent se- mantic analysis,â Machine learning, vol. 42, no. 1-2, pp. 177â196, 2001.
[92] D. M. Blei, A. Y. Ng, and M. I. Jordan, âLatent dirichlet alloca- tion,â Journal of Machine Learning Research, vol. 3, pp. 993â1022, 2003.
[93] E. H¨orster, R. Lienhart, and M. Slaney, âImage retrieval on large- scale image databases,â in ACM International Conference on Image and Video Retrieval, 2007, pp. 17â24.
[94] R. Lienhart and M. Slaney, âpLSA on large scale image databases,â in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 4, 2007, pp. IVâ1217.
19
[95] K. Simonyan and A. Zisserman, âVery deep convolutional networks for large-scale image recognition,â arXiv preprint arXiv:1409.1556, 2014. | 1706.06064#126 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 127 | [96] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, âGoing deeper with convolutions,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
97] 98, Y. Bengio, âLearning deep architectures for ai,â Foundations and trends® in Machine Learning, vol. 2, no. 1, pp. 1-127, 2009. E. Hérster and R. Lienhart, âDeep networks for image retrieval on large-scale databases,â in ACM International Conference on Multimedia. ACM, 2008, pp. 643-646. | 1706.06064#127 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 128 | [99] A. Krizhevsky, I. Sutskever, and G. E. Hinton, âImagenet classiï¬- cation with deep convolutional neural networks,â in Advances in Neural Information Processing Systems (NIPS), 2012, pp. 1097â1105. [100] A. Sharif Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, âCNN features off-the-shelf: an astounding baseline for recogni- tion,â in IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2014.
[101] J. Wan, D. Wang, S. C. H. Hoi, P. Wu, J. Zhu, Y. Zhang, and J. Li, âDeep learning for content-based image retrieval: A comprehen- sive study,â in ACM International Conference on Multimedia (MM). ACM, 2014, pp. 157â166.
[102] A. S. Razavian, J. Sullivan, A. Maki, and S. Carlsson, âVisual instance retrieval with deep convolutional networks,â arXiv preprint arXiv:1412.6574, 2014. | 1706.06064#128 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 130 | [105] J. R. Uijlings, K. E. van de Sande, T. Gevers, and A. W. Smeulders, âSelective search for object recognition,â International Journal of Computer Vision (IJCV), vol. 104, no. 2, pp. 154â171, 2013. [106] B. Alexe, T. Deselaers, and V. Ferrari, âMeasuring the objectness of image windows,â IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 34, no. 11, pp. 2189â2202, 2012. [107] M.-M. Cheng, Z. Zhang, W.-Y. Lin, and P. Torr, âBing: Binarized normed gradients for objectness estimation at 300fps,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.
[108] S. Sun, W. Zhou, Q. Tian, and H. Li, âScalable object retrieval with compact image representation from generic object regions,â ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), vol. 12, no. 2, p. 29, 2015. | 1706.06064#130 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 131 | [109] G. Tolias, R. Sicre, and H. J´egou, âParticular object retrieval with integral max-pooling of cnn activations,â International Conference on Learning and Representation (ICLR), 2016.
[110] A. Gordo, J. Almazan, J. Revaud, and D. Larlus, âDeep image retrieval: Learning global representations for image search,â in European Conference on Computer Vision (ECCV), 2016.
[111] S. Ren, K. He, R. Girshick, and J. Sun, âFaster r-cnn: Towards real-time object detection with region proposal networks,â in Advances in Neural Information Processing Systems (NIPS), 2015, pp. 91â99.
[112] A. Babenko, A. Slesarev, A. Chigorin, and V. Lempitsky, âNeural codes for image retrieval,â in European Conference on Computer Vision (ECCV). Springer, 2014, pp. 584â599. | 1706.06064#131 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 132 | [113] M. Paulin, M. Douze, Z. Harchaoui, J. Mairal, F. Perronin, and C. Schmid, âLocal convolutional features with unsupervised training for image retrieval,â in IEEE International Conference on Computer Vision (ICCV), 2015, pp. 91â99.
[114] R. Xia, Y. Pan, H. Lai, C. Liu, and S. Yan, âSupervised hashing for image retrieval via image representation learning,â in AAAI Conference on Artiï¬cial Intellignece, 2014, pp. 2156â2162.
[115] H. Lai, Y. Pan, Y. Liu, and S. Yan, âSimultaneous feature learning and hash coding with deep neural networks,â arXiv preprint arXiv:1504.03410, 2015.
[116] H. J´egou, M. Douze, C. Schmid, and P. P´erez, âAggregating local descriptors into a compact image representation,â in IEEE Conference on Computer Vision and Pattern Recognition, 2010, pp. 3304â3311. | 1706.06064#132 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 133 | [117] F. Perronnin, Y. Liu, J. S´anchez, and H. Poirier, âLarge-scale image retrieval with compressed ï¬sher vectors,â in IEEE Conference on
Computer Vision and Pattern Recognition (CVPR). 3384â3391. IEEE, 2010, pp.
[118] F. Li, W. Tong, R. Jin, A. K. Jain, and J.-E. Lee, âAn efï¬cient key point quantization algorithm for large scale image retrieval,â in ACM workshop on Large-scale Multimedia Retrieval and Mining. ACM, 2009, pp. 89â96.
[119] L. Chu, S. Wang, Y. Zhang, S. Jiang, and Q. Huang, âGraph- density-based visual word vocabulary for image retrieval,â in IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2014, pp. 1â6. | 1706.06064#133 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 134 | [120] W. Dong, Z. Wang, M. Charikar, and K. Li, âEfï¬ciently matching sets of features with random histograms,â in ACM International Conference on Multimedia (MM). ACM, 2008, pp. 179â188. [121] W. Zhou, M. Yang, H. Li, X. Wang, Y. Lin, and Q. Tian, âTo- wards codebook-free: Scalable cascaded hashing for mobile im- age search,â IEEE Transactions on Multimedia, vol. 16, no. 3, pp. 601â611, 2014.
[122] S. Zhang, Q. Tian, G. Hua, Q. Huang, and W. Gao, âGenerating descriptive visual words and visual phrases for large-scale image applications,â IEEE Transactions on Image Processing (TIP), vol. 20, no. 9, pp. 2664â2677, 2011.
[123] X. Wang, M. Yang, T. Cour, S. Zhu, K. Yu, and T. X. Han, âCon- textual weighting for vocabulary tree based image retrieval,â in International Conference on Computer Vision, 2011, pp. 209â216. | 1706.06064#134 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 135 | [124] Z. Liu, H. Li, W. Zhou, and Q. Tian, âEmbedding spatial context information into inverted ï¬le for large-scale image retrieval,â in ACM International Conference on Multimedia, 2012, pp. 199â208.
[125] Z. Liu, H. Li, W. Zhou, R. Zhao, and Q. Tian, âContextual hashing for large-scale image search,â IEEE Transactions on Image Processing (TIP), vol. 23, no. 4, pp. 1606â1614, 2014.
[126] O. Chum, M. Perdoch, and J. Matas, âGeometric min-hashing: Finding a (thick) needle in a haystack,â in IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 17â24.
[127] D. N. Bhat and S. K. Nayar, âOrdinal measures for image cor- respondence,â IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 20, no. 4, pp. 415â423, 1998. | 1706.06064#135 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 136 | [128] S. Lazebnik, C. Schmid, and J. Ponce, âBeyond bags of fea- tures: Spatial pyramid matching for recognizing natural scene categories,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol. 2.
[129] Y. Cao, C. Wang, Z. Li, L. Zhang, and L. Zhang, âSpatial-bag- of-features,â in IEEE Conference on Computer Vision and Pattern Recognition, 2010, pp. 3352â3359.
[130] Z. Wu, Q. Ke, J. Sun, and H.-Y. Shum, âScalable face image retrieval with identity-based quantization and multireference reranking,â IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 10, pp. 1991â2001, 2011.
[131] J. L. Bentley, âK-d trees for semidynamic point sets,â in Annual Symp. Computational Geometry, 1990, pp. 187â197.
[132] C. Silpa-Anan and R. Hartley, âLocalization using an image map,â in Australian Conference on Robotics and Automation, 2004. | 1706.06064#136 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 137 | [132] C. Silpa-Anan and R. Hartley, âLocalization using an image map,â in Australian Conference on Robotics and Automation, 2004.
[133] M. Muja and D. G. Lowe, âScalable nearest neighbor algorithms for high dimensional data,â IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 36, 2014.
[134] W. Zhou, M. Yang, X. Wang, H. Li, Y. Lin, and Q. Tian, âScalable feature matching by dual cascaded scalar quantization for im- age retrieval,â IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 38, no. 1, pp. 159â171, 2016.
[135] M. Jain, H. J´egou, and P. Gros, âAsymmetric hamming embed- ding: taking the best of our bits for large scale image search,â in ACM International Conference on Multimedia, 2011, pp. 1441â1444. [136] W. Zhou, H. Li, Y. Lu, M. Wang, and Q. Tian, âVisual word expansion and BSIFT veriï¬cation for large-scale image search,â Multimedia Systems, vol. 21, no. 3, pp. 245â254, 2013. | 1706.06064#137 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 138 | [137] Y. Xia, K. He, F. Wen, and J. Sun, âJoint inverted indexing,â in International Conference on Computer Vision, 2013.
[138] H. J´egou and O. Chum, âNegative evidences and co-occurences in image retrieval: The beneï¬t of PCA and whitening,â in Euro- pean Conference on Computer Vision, 2012, pp. 774â787.
[139] L. Zheng, S. Wang, W. Zhou, and Q. Tian, âBayes merging of multiple vocabularies for scalable image retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition, 2014. [140] P. Indyk and R. Motwani, âApproximate nearest neighbors: to- wards removing the curse of dimensionality,â in Annual ACM Symposium Theory of Computing. ACM, 1998, pp. 604â613.
20
[141] A. Andoni and P. Indyk, âNear-optimal hashing algorithms for approximate nearest neighbor in high dimensions,â in IEEE Symposium Foundations of Computer Science, 2006, pp. 459â468. | 1706.06064#138 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 139 | [142] Q. Lv, W. Josephson, Z. Wang, M. Charikar, and K. Li, âMulti- probe lsh: efï¬cient indexing for high-dimensional similarity search,â in International Conference Very Large Data Bases, 2007, pp. 950â961.
[143] J. Wang, S. Kumar, and S.-F. Chang, âSemi-supervised hashing for scalable image retrieval,â in IEEE Conference on Computer Vision and Pattern Recognition, 2010, pp. 3424â3431.
[144] Y. Gong and S. Lazebnik, âIterative quantization: A procrustean approach to learning binary codes,â in IEEE Conference on Com- puter Vision and Pattern Recognition, 2011, pp. 817â824.
[145] D. Aiger, E. Kokiopoulou, and E. Rivlin, âRandom grids: Fast approximate nearest neighbors and range searching for image search,â in International Conference on Computer Vision, 2013. [146] M. Iwamura, T. Sato, and K. Kise, âWhat is the most efï¬cient way to select nearest neighbor candidates for fast approximate nearest neighbor search?â in International Conference on Computer Vision, 2013. | 1706.06064#139 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 140 | [147] J. Wang and S. Li, âQuery-driven iterated neighborhood graph search for large scale indexing,â in ACM International Conference on Multimedia (MM). ACM, 2012, pp. 179â188.
[148] M. Wang, W. Zhou, Q. Tian, Z. Zha, and H. Li, âLinear distance preserving pseudo-supervised and unsupervised hashing,â in ACM International Conference on Multimedia (MM). ACM, 2016, pp. 1257â1266.
[149] T. Ge, K. He, Q. Ke, and J. Sun, âOptimized product quantization for approximate nearest neighbor search,â in IEEE Conference on Computer Vision and Pattern Recognition, 2013.
[150] T. Tuytelaars and C. Schmid, âVector quantizing feature space with a regular lattice,â in International Conference on Computer Vision, 2007, pp. 1â8.
[151] R. Arandjelovic and A. Zisserman, âAll about vlad,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2013, pp. 1578â1585. | 1706.06064#140 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 141 | I. Kompatsiaris, G. Tsoumakas, and I. Vlahavas, âA comprehensive study over vlad and product quantizationin for large-scale image retrieval,â IEEE Transactions on Multimedia (TMM), 2014.
[153] H. J´egou and A. Zisserman, âTriangulation embedding and democratic aggregation for image search,â in IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2014, pp. 3310â3317.
[154] Z. Gao, J. Xue, W. Zhou, S. Pang, and Q. Tian, âFast democratic aggregation and query fusion for image search,â in ACM Interna- tional Conference on Multimedia Retrieval (ICMR), 2015.
[155] T. Ge, Q. Ke, and J. Sun, âSparse-coded features for image retrieval.â British Machine Vision Conference (BMVC), 2013.
[156] Z. Liu, H. Li, W. Zhou, T. Rui, and Q. Tian, âUniforming residual vector distribution for distinctive image representation,â IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2015. | 1706.06064#141 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 142 | [157] Z. Liu, H. Li, W. Zhou, and Q. Tian, âUniting keypoints: Local visual information fusion for large scale image search,â IEEE Transactions on Multimedia (TMM), 2015.
[158] T. Jaakkola and D. Haussler, âExploring generative model in discriminative classiï¬ers,â in Advances in Neural Information Pro- cessing Systems (NIPS), 1998.
[159] T. Jaakkola, D. Haussler et al., âExploiting generative models in discriminative classiï¬ers,â Advances in Neural Information Process- ing Systems (NIPS), pp. 487â493, 1999. | 1706.06064#142 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
1706.06064 | 143 | [160] J. S´anchez, F. Perronnin, T. Mensink, and J. Verbeek, âImage clas- siï¬cation with the ï¬sher vector: theory and practice,â International Journal of Computer Vision (IJCV), vol. 105, no. 3, pp. 222â245, 2013. [161] L.-Y. Duan, F. Gao, J. Chen, J. Lin, and T. Huang, âCompact descriptors for mobile visual search and mpeg cdvs standard- ization,â in IEEE International Symposium on Circuits and Systems (ISCAS).
[162] Y. Gong, L. Wang, R. Guo, and S. Lazebnik, âMulti-scale orderless pooling of deep convolutional activation features,â in European Conference on Computer Vision (ECCV). Springer, 2014, pp. 392â 407.
[163] A. Babenko and V. Lempitsky, âAggregating local deep features for image retrieval,â in IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1269â1277.
[164] R. Baeza-Yates, B. Ribeiro-Neto et al., Modern information retrieval. ACM press New York., 1999, vol. 463. | 1706.06064#143 | Recent Advance in Content-based Image Retrieval: A Literature Survey | The explosive increase and ubiquitous accessibility of visual data on the Web
have led to the prosperity of research activity in image search or retrieval.
With the ignorance of visual content as a ranking clue, methods with text
search techniques for visual retrieval may suffer inconsistency between the
text words and visual content. Content-based image retrieval (CBIR), which
makes use of the representation of visual content to identify relevant images,
has attracted sustained attention in recent two decades. Such a problem is
challenging due to the intention gap and the semantic gap problems. Numerous
techniques have been developed for content-based image retrieval in the last
decade. The purpose of this paper is to categorize and evaluate those
algorithms proposed during the period of 2003 to 2016. We conclude with several
promising directions for future research. | http://arxiv.org/pdf/1706.06064 | Wengang Zhou, Houqiang Li, Qi Tian | cs.MM, cs.IR | 22 pages | null | cs.MM | 20170619 | 20170902 | [
{
"id": "1504.03410"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.