image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 8:Qualitative comparison of methods for USV-based Obstacle Segmentation. Prominent errors and examples of good behaviour are highlighted in red and cyan, respectively. | 2501.10343v1 | 16 |
|
Figure 9:Comparison of F1 detection score for the best3333submissions from Section3.3under various environmental conditions (environment type, reflections, and conditions). | 2501.10343v1 | 17 |
|
Figure 10:Qualitative comparison of the best3333submissions for USV-based embedded obstacle segmentation from Section3.3. | 2501.10343v1 | 18 |
|
Figure 11:USV-based Panoptic Segmentation: Performance of top methods in terms of panoptic quality (PQ) w.r.t. different scene attributes. The performance of the baseline method (Mask2Former[13]w/ Swin-B) is marked as a dotted line for reference. | 2501.10343v1 | 19 |
|
Figure 12:USV-based Panoptic Segmentation: Dynamic obstacle detection rate of top methods with respect to ground-truth obstacle size. | 2501.10343v1 | 20 |
|
Figure 14:Qualitative comparison of underwater image restoration proposed by top performing teams in MarineVision Restoration Challenge. | 2501.10343v1 | 22 |
|
Figure 16:Overview of the architecture proposed by Team NJUST-KMG for underwater image restoration and marine species detection. | 2501.10343v1 | 24 |
|
Figure 17:Overview of the architecture proposed by Team BUPT MCPRL for underwater image restoration and marine species detection. | 2501.10343v1 | 25 |
|
Figure 6:Regularization techniques were applied in the proposed model to prevent overfitting and improve performance. | 2501.10342v1 | 6 |
|
Figure 4:Robustness of PASTIS.(a) Exact match accuracy with and without trapezoid modification for large time intervalsΔtΔ𝑡\Delta troman_Δ italic_t, for the Lorenz model with long trajectories (τ=4×104𝜏4superscript104\tau=4\times 10^{4}italic_τ = 4 × 10 start_POSTSUPERSCRIPT 4 end_POSTSUPERSCRIPT). (b) Same
for the Stratonovich modification for measurement noise. Herext→xt+ηabsent→subscript𝑥𝑡subscript𝑥𝑡𝜂x_{t}\xrightarrow{}x_{t}+\etaitalic_x start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_ARROW start_OVERACCENT end_OVERACCENT → end_ARROW italic_x start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT + italic_ηwhereη∼𝒩(0,σ2)similar-to𝜂𝒩0superscript𝜎2\eta\sim\mathcal{N}(0,\sigma^{2})italic_η ∼ caligraphic_N ( 0 , italic_σ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ). | 2501.10339v2 | 4 |
|
Figure 2:(a) Comparison between the success rate distributions of ground-truth and agent-simulated response data. (b) Using LLMs as judges to identify whether the records of the agent’s simulations originate from real humans. (c) The ablation studies. | 2501.10332v1 | 2 |
|
Figure 1:a) A standard forecasting setup, where an evergreen item has past observations to exploit,e.g., # sales.
b)NFPPFproblem, where no past observations are available, and exogenous data must be considered.
The figure is from[18]. | 2501.10324v1 | 1 |
|
Figure 2:Examples of data in the VISUELLE 2.0 dataset.
The figure reports (from left to right) the product’s image, textual tags, sales time series, restocking information, inventory time series, and discount time series.
Figures are from[23]. | 2501.10324v1 | 2 |
|
Figure 3:PRISMA-based flowchart of the retrieval process.
Due to space issues, the diagram has been reformatted into a horizontal rather than the standard vertical format. | 2501.10324v1 | 3 |
|
Figure 2:Bytes per word (BPW) and bytes per token (BPT) statistics, showing that words are a coarser unit than subword tokens.
The tokenizer has been fitted to the DCLM-baseline dataset (left).
On a dataset to which the tokenizer is not attuned, such as the German Occiglot dataset (right), the tokenizer “fragments” and BPT drops significantly. | 2501.10322v2 | 2 |
|
Figure 3:Word and byte accuracy for hierarchical models with different encoder/decoder sizes. Each candidate encoder/decoder size has been trained with backbone sizes ranging fromLb=Hb=16subscript𝐿bsubscript𝐻b16L_{\text{b}}=H_{\text{b}}=16italic_L start_POSTSUBSCRIPT b end_POSTSUBSCRIPT = italic_H start_POSTSUBSCRIPT b end_POSTSUBSCRIPT = 16toLb=Hb=30subscript𝐿bsubscript𝐻b30L_{\text{b}}=H_{\text{b}}=30italic_L start_POSTSUBSCRIPT b end_POSTSUBSCRIPT = italic_H start_POSTSUBSCRIPT b end_POSTSUBSCRIPT = 30. The horizontal axis shows training compute (Eq.7) normalized by the highest value.
The vertical gray lines indicate the compute required by the baseline models.
Small character-level modules tend to yield better word accuracy at the tested compute budgets. | 2501.10322v2 | 3 |
|
Figure 4:Average change in accuracy across the MMLU, OpenBookQA, Arc Challenge and HellaSwag eval tasks for perturbations to the prompt. | 2501.10322v2 | 4 |
|
Figure 5:Continued pretraining on Occiglot German. We show average eval scores on English and German.
The hierarchical architecture adapts better to the new language, while also retaining higher scores on English evals.
At an equal FLOP budget we are only able to train the baseline for half as many steps due to it requiring almost two times the compute per document on the german dataset. Nevertheless, we continued training the baseline for more steps, depicted as dotted lines. | 2501.10322v2 | 5 |
|
Figure 6:Scale up of the character-level module size at different aspect ratios and a fixed backbone size of 16 heads and layers. The x-axis depicts compute cost (Eq.7), normalized by the largest value. | 2501.10322v2 | 6 |
|
Figure 6:Scale up of the character-level module size at different aspect ratios and a fixed backbone size of 16 heads and layers. The x-axis depicts compute cost (Eq.7), normalized by the largest value. | 2501.10322v2 | 7 |
|
Figure 7:Encoder-decoder balance.
A fixed number of8888character-level layers is distributed between encoder and decoder.
The backbone is fixed. | 2501.10322v2 | 8 |
|
Figure 9:Learning curve in bits per byte (bpb) for continued pretraining on the Chinese Skypile dataset.
Using the new Unicode splitter, the hierarchical architecture adapts to the new language more rapidly and achieves better bpb within the assigned step budget. Note that the baseline requires 2.3 times more compute for the same number of training steps. | 2501.10322v2 | 10 |
|
Figure 9:Learning curve in bits per byte (bpb) for continued pretraining on the Chinese Skypile dataset.
Using the new Unicode splitter, the hierarchical architecture adapts to the new language more rapidly and achieves better bpb within the assigned step budget. Note that the baseline requires 2.3 times more compute for the same number of training steps. | 2501.10322v2 | 11 |
|
Figure 10:Learning curve in bits per byte (bpb) for continued pretraining on the Chinese Skypile dataset.
Both variants use the Unicode splitter for the continued pretraining stage, but have been trained using different splitters in the initial English-only pretraining phase.
The variant that undergoes a change of splitting rule adapts quickly and achieves the same final performance. | 2501.10322v2 | 12 |
|
Figure 1:Comparison of performance and computational cost of the language decoder between the original and HiMix models.The circles, arranged from smallest to largest, represent the models Qwen2-0.5B, Llama3-1B, TinyLlama-1.1B, and Llama3-3B. While maintaining a similar performance to the original models, our HiMix achieves a10×10\times10 ×reduction in computational cost. | 2501.10318v1 | 1 |
|
Figure 3:Comparison of Vanilla Model and HiMix Architectures.Left:Overall structure of traditional Vanilla.Middle:Overall structure of HiMix.Right:Details of HiMix. | 2501.10318v1 | 3 |
|
Figure 4:Exploration of Model Architecture Design. (a) Uniform Vision Injection to Each Layer. (b) Hierarchical Vision Injection Through Multi-Level Connectors. The Mixture Attention (MA) differs from the main method in that the vision and language sequences share the KV projection layers | 2501.10318v1 | 4 |
|
Figure 5:Performance gaps between HiMix and the baseline model across various benchmarks under two training strategies:Regular ParadigmandEnhanced Paradigm. The y-axis represents the performance gap relative to the baseline, with positive values indicating improvements. Subfigures (a), (b), and (c) show results for Qwen2-0.5B, Llama3-1B, and TinyLLaMA-1.1B, respectively. Results indicate that using the Enhanced Paradigm amplifies HiMix’s performance benefits over the baseline. | 2501.10318v1 | 5 |
|
Figure 7:Qualitative results of yes/no questions.Correct answersandWrong answersare highlighted in color respectively. | 2501.10318v1 | 7 |
|
Figure 8:Qualitative comparison of image captions.Correct descriptionsandhallucinatedare highlighted in color respectively. | 2501.10318v1 | 8 |
|
Figure 9:Qualitative comparison of image captions.Correct descriptionsandhallucinatedare highlighted in color respectively. | 2501.10318v1 | 9 |
|
Figure 10:Qualitative comparison of image captions.Correct descriptionsandhallucinatedare highlighted in color respectively. | 2501.10318v1 | 10 |
|
Figure 11:Qualitative results of object recognition tasks.Correctly identified objectsanderrorsare highlighted in color respectively. | 2501.10318v1 | 11 |
|
Figure 12:Qualitative results of text OCR tasks.Correctly recognized textanderrorsare highlighted for comparison. | 2501.10318v1 | 12 |
|
Figure 3:Performance gain (JGA) with varyingτfpsubscript𝜏fp\tau_{\operatorname{fp}}italic_τ start_POSTSUBSCRIPT roman_fp end_POSTSUBSCRIPTandτfnsubscript𝜏fn\tau_{\operatorname{fn}}italic_τ start_POSTSUBSCRIPT roman_fn end_POSTSUBSCRIPTon MultiWOZ test set with Llama. Thex𝑥xitalic_x-axis shows the percentage of turns that are impacted. On average, 1.1 slots have been involved in the corrections consistently over the range of this plot. Hence, we did not include it in the depiction. | 2501.10316v1 | 3 |
|
Figure 2:Taking multi-view captures as input, GSTAR tracks and reconstructs dynamic objects frame by frame. For each frame, GSTAR first warps the previous frame’s result using scene flow (Sec.3.2). It then reconstructs Gaussian Surfaces (Gaussian-attached mesh,Sec.3.1) by fixed-topology reconstruction (Sec.3.3). To handle topology-changing surfaces, GSTAR detects topology changes, unbinds Gaussians on these surfaces, and adds new Gaussians as needed (Sec.3.4). Finally, the Gaussian Surfaces are updated through re-meshing (Sec.3.5). | 2501.10283v2 | 2 |
|
Figure 3:Details of the mesh update process. (a) Visualization of unbinding weights defined inEq.10, where red indicates high weights in topology-changing regions.
(b) Mesh connection process between original and new surfaces, with blue dotted lines showing vertex correspondences. | 2501.10283v2 | 3 |
|
Figure 4:Comparisons of appearance and geometry reconstruction. Dynamic 3D Gaussians[26]and PhysAvatar[50]yield suboptimal reconstruction results. HumanRF[17]and 2DGS[16], lacking tracking capabilities, struggle under heavy occlusion. In contrast, GSTAR provides high-quality reconstruction while supporting tracking. Additional comparisons are provided in our supplementary materials. | 2501.10283v2 | 4 |
|
Figure 5:Tracking comparisons using AprilTags. GSTAR achieves more accurate tracking results, with predicted (red) and ground truth (blue) trajectories of tag centers shown. | 2501.10283v2 | 5 |
|
Figure 6:Applications.
(a) Object editing: virtual objects sync with dynamic surfaces.
(b) Appearance editing: texture changes propagate across frames.
(c) General applicability to diverse scenes including multiple people, human-object interactions, and robotic motion. | 2501.10283v2 | 6 |
|
Figure 5:Visualization of the detection on three methods. Ground truths are marked in green, while predicted boxes are red. | 2501.10266v1 | 5 |
|
Figure 2:Segmented waveform of speaker M02 pronouncing the sentence “Carl lives in a lively home”. Ground truth phonemic transcriptions are shown at the bottom for reference (’noi’ corresponds to noise). | 2501.10256v1 | 2 |
|
Figure 4:WER results on Torgo, grouped by severity level, presented for each experimental configuration. | 2501.10256v1 | 8 |
|
Figure 1:System model for the multi-sensor wireless inference with analog memristor neural networks. | 2501.10245v1 | 1 |
|
Figure 3:Inference accuracy for the ModelNet dataset with a multi-modal system withM=4𝑀4M=4italic_M = 4digital sensors each having different computational capabilities. | 2501.10245v1 | 3 |
|
Figure 4:Inference accuracy for learnable sensor fusion using CARLA dataset withM=5𝑀5M=5italic_M = 5digital sensors with perfect (fixed) SNR training. | 2501.10245v1 | 4 |
|
Figure 5:Inference accuracy for learnable sensor fusion using CARLA dataset withM=5𝑀5M=5italic_M = 5digital sensors with SNR-robust training. | 2501.10245v1 | 5 |
|
Figure 6:Test accuracy for MNIST classification with digital and analog sensors against number of sensors. | 2501.10245v1 | 6 |
|
Figure 7:Test accuracy for ModelNet classification utilizing digital and analog sensors, with identical SNR during training. | 2501.10245v1 | 7 |
|
Figure 8:Test accuracy for ModelNet classification utilizing digital and analog sensors, with 10dB SNR during training. | 2501.10245v1 | 8 |
|
Figure 9:Test accuracy for ModelNet classification utilizing digital and analog sensors, with a random SNR during training. | 2501.10245v1 | 9 |
|
Figure 4:Illustrative example of the simplex polyhedron and the five moves
of the Nelder-Mead Local Search. Source:Chaves et al. (2024). | 2501.10243v2 | 4 |
|
Figure 10:Example of a solution found by BRKGA-QL for a literature instance bounded by bed availability. | 2501.10243v2 | 10 |
|
Figure 11:Example of a solution found by BRKGA-QL for a case study instance bounded by OR availability. | 2501.10243v2 | 11 |
|
Figure 1:Data flow for model building pipeline (A) and model implementation (B). Two databases are exemplified as data sources, EHR and ICU, although multiple other sources might be used in the hospital’s flow and for data extraction. EHR = Electronic Health Record; DB = database; ICU = Intensive Care Unit; REST API = RESTful application programming interface. | 2501.10240v1 | 1 |
|
Figure 1:Illustration of the counterfactuals generated (green and yellow points) for the given factual y (marked as red cross) under the assumption of Euclidean distance and different feasibility and plausibility constraints. | 2501.10234v1 | 1 |
|
Figure 2:Illustration of the counterfactuals computed in the case of a 2-d synthetic dataset partitioned in two clusters usingk𝑘kitalic_k-means. The factualy𝑦yitalic_yis the same and the counterfactuals computed for several values of maskM𝑀Mitalic_Mand palusibility factorϵitalic-ϵ\epsilonitalic_ϵare presented. | 2501.10234v1 | 2 |
|
Figure 3:Illustration of the counterfactuals computed in the case of two Gaussian clusters with full covariances. The factualy𝑦yitalic_yis the same and the counterfactuals computed for several values of maskM𝑀Mitalic_Mand plausibility factorϵitalic-ϵ\epsilonitalic_ϵare presented. | 2501.10234v1 | 3 |
|
Figure 5:k𝑘kitalic_k-means for 2D: (left) no immutable features, (middle)-(right) one immutable feature. | 2501.10234v1 | 16 |
|
Figure 6:k𝑘kitalic_k-means for 3D: (left) no immutable features, (middle) one immutable feature,(right) two immutable feature. | 2501.10234v1 | 20 |
|
Figure 7:k𝑘kitalic_k-means for Iris: (left) no immutable features, (middle) one immutable feature, (right) two immutable features. | 2501.10234v1 | 24 |
|
Figure 8:k𝑘kitalic_k-means for Wine: (left) no immutable features, (middle) four immutable feature, (right) seven immutable features. | 2501.10234v1 | 28 |
|
Figure 9:k𝑘kitalic_k-means for Pendigits: (left) no immutable features, (middle) three immutable feature, (left) six immutable features. | 2501.10234v1 | 32 |
|
Figure 10:Gaussian clustering for 2D: (left) no immutable features, (middle)-(right) one immutable feature. | 2501.10234v1 | 36 |
|
Figure 11:Gaussian clustering for 3D: (left) no immutable features, (middle) one immutable feature, (right) two immutable features. | 2501.10234v1 | 40 |
|
Figure 12:Gaussian clustering for Iris: (left) no immutable features, (middle) one immutable feature, (right) two immutable features. | 2501.10234v1 | 44 |
|
Figure 13:Gaussian clustering for Wine: (left) No immutable features, (middle) four immutable feature, (right) seven immutable features. | 2501.10234v1 | 48 |
|
Figure 14:Gaussian for Pendigits: (left) no immutable features, (middle) three immutable feature, (right) six immutable features. | 2501.10234v1 | 52 |
|
Figure 1:Illustration of a rigid body transformation at two distinct locations𝑺(0)superscript𝑺0\bm{S}^{(0)}bold_italic_S start_POSTSUPERSCRIPT ( 0 ) end_POSTSUPERSCRIPTand𝑺(1)superscript𝑺1\bm{S}^{(1)}bold_italic_S start_POSTSUPERSCRIPT ( 1 ) end_POSTSUPERSCRIPT. Without loss of generality, we set the initial position to be identical to the matrix𝑪𝑪\bm{C}bold_italic_C, which defines the shape and orientation of the rigid body. The second location𝑺(1)superscript𝑺1\bm{S}^{(1)}bold_italic_S start_POSTSUPERSCRIPT ( 1 ) end_POSTSUPERSCRIPTof the body is then determined according to equation (1), and is obtained by the transformation of𝑺(0)superscript𝑺0\bm{S}^{(0)}bold_italic_S start_POSTSUPERSCRIPT ( 0 ) end_POSTSUPERSCRIPTvia a rotation matrix𝑸𝑸\bm{Q}bold_italic_Qand a translation vector𝒕𝒕\bm{t}bold_italic_t. | 2501.10219v1 | 1 |
|
Figure 3:RMSEof the translation estimate of thegenie-aided(GA)proposed methods and theSotA, over the range errorσ𝜎\sigmaitalic_σ. | 2501.10219v1 | 2 |
|
Figure 4:RMSEof the translation estimate of the proposedMDS-based method for different levels of available information, over the range errorσ𝜎\sigmaitalic_σ. | 2501.10219v1 | 3 |
|
Figure 5:RMSEof the translation estimate of the proposed method aided by matrix completion for different levels of available information, over the range errorσ𝜎\sigmaitalic_σ. | 2501.10219v1 | 4 |
|
Figure 6:RMSEof the translation estimate of the proposed robust method compared to theMDS-based method aided by matrix completion for different levels of available information, over the range errorσ𝜎\sigmaitalic_σ. | 2501.10219v1 | 5 |
|
Figure 7:Illustration of a rigid body modeled as a unit vector𝒗Psubscript𝒗𝑃\bm{v}_{P}bold_italic_v start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT, with the true transformed vector𝒗Tsubscript𝒗𝑇\bm{v}_{T}bold_italic_v start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPTand its estimate𝒗^Tsubscript^𝒗𝑇\hat{\bm{v}}_{T}over^ start_ARG bold_italic_v end_ARG start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPT. | 2501.10219v1 | 6 |
|
Figure 8:RMSEof the genie-aided orientation estimate of the proposed methods in combination with theSotA, compared to the pureSotAsolution for different levels of available information, over the range errorσ𝜎\sigmaitalic_σ. | 2501.10219v1 | 7 |
|
Figure 9:RMSEof the egoistic orientation estimate of the combinations of the proposed methods for different levels of available information, over the range errorσ𝜎\sigmaitalic_σ. | 2501.10219v1 | 8 |
|
Fig. 1:High-level depiction of Chronos: Transform input time series into tokens via scaling an quantization; train the language model via cross-entropy loss; autoregressively sample multiple trajectories, then map to numerical values to obtain predictive distribution. | 2501.10216v1 | 1 |
|
Figure 2:The overall training pipeline of Disharmony is depicted here. The training process is divided into two stages: initially, we pretrain MaskFormer using the DISK25k dataset(Step A). Subsequently, we fine-tune this model with the IH Dataset(Step B) and the RealHM Dataset(Step C), culminating in the development of Disharmony. | 2501.10212v1 | 2 |
|
Figure 3:Samples from the test set. We randomly selected 100 images from the iHarmony4 dataset[7], each consisting of a ground truth image, a composite image (randomly chosen one composite from different composites), and a corresponding mask. The composite images and masks were then processed using various harmonization methods (DoveNet[8], Harmonizer[16], HT[13], Hi-Net[5], and PCT-Net[11]) to generate the test images. | 2501.10212v1 | 3 |
|
Figure 4:Qualitative results of Disharmony across different harmonization methods(DoveNet[8], Harmonizer[16], HT[13], Hi-Net[5], and PCT-Net[11]) and composite images. For clarity, the background is shown in green to make the segmentation masks more visually discernible, with the resulting segmentation mask retaining the original colors of the image. | 2501.10212v1 | 4 |
|
Figure 5:Qualitative comparison of various forensic methods(MantraNet[28], Cat-Net[18], HiFi-Net[12], IML-ViT[20]) and Disharmony across different harmonization methods(DoveNet[8], Harmonizer[16], HT[13], Hi-Net[5], and PCT-Net[11]), using the given image and mask. | 2501.10212v1 | 5 |
|
Figure 6:Evaluation of the aggregation method. We tested three different scenarios: (1) pretraining MaskFormer[6]with DISK25K[24], (2) pretraining with DISK25K followed by fine-tuning with the RealHM dataset[14], and (3) pretraining with DISK25K followed by fine-tuning with both the RealHM and IH dataset[4]. Based on these evaluations, we concluded that the most effective approach is to pretrain with DISK25K and fine-tune with both the RealHM and IH dataset. | 2501.10212v1 | 6 |
|
Figure 7:The resulting masks produced by various forensic models(MantraNet[28], Cat-Net[18], HiFi-Net[12], IML-ViT[20]) and Disharmony, given an unaltered input image | 2501.10212v1 | 7 |
|
Figure 8:The resulting images produced by text edits using InstructPix2Pix[3]and Imagic[15]on the TEDBench dataset. Note that the backgrounds change even when the text instructions do not specify modifications to the background. | 2501.10212v1 | 8 |
|
Figure 10:The resulting segmentation outcomes for images that have been edited using DragDiffusion[23] | 2501.10212v1 | 10 |
|
Figure 1:Relationship between hyperparameterk𝑘kitalic_kand FPR95 on CIFAR-10 (left) and CIFAR-100 (right) for ResNet-34 Supervised Contrastive classifier features. The blue line shows HACk-OOD for fixedk𝑘kitalic_kvalues. The orange line representsAdaptivek𝑘kitalic_kwith differentk𝑘kitalic_kvalues per label and regularization. The green line showsAdaptivek𝑘kitalic_kwithout regularization. | 2501.10209v1 | 1 |
|
Figure 3:HACk-OOD steps to generate a class contour in two dimensions. It illustrates what a single ID cluster’s contour would look like in two dimensional space. Data was generated by drawing 5000 observations sampled from 5 Gaussian distributions and placing the cluster means sufficiently close to represent one larger cluster which varies non-uniformly. Sub-figures (a)-(f) show a representation of how one hypercone is constructed, sub-figure (g) shows a model representation of what the shape of the expected contour would be. Sub-figure (h) shows in blue the actual shape of HACk-OOD’s contour when running HACk-OOD on the this synthetic data. | 2501.10209v1 | 3 |
|
Figure 1:Illustration of the proposed real-time processing technique for the data of a hyperspectral push-broom camera for on-board phenotype segmentation of trees. | 2501.10199v1 | 1 |
|
Figure 2:Distributions of the PROSPECT-D model parameters for the leaves in the generated data. This is to give an idea of the distribution of the labels in the dataset and of what values the different parameters can take. These are the values for chlorophyll, carotenoid, brown content, water thickness, dry matter, and anthocyanin respectively. | 2501.10199v1 | 2 |
|
Figure 3:Example simulation data: fake RGB image on the top left, with spectra inside the highlighted squares on the top right. Best viewed in colour. The bottom to panel show the pixel labels for the trees. These contain values of 0 or 1 for if the tree is of certain class or not. | 2501.10199v1 | 3 |
|
Figure 4:Illustrative figure of the model and the pipeline. OHSLIC takes a line as input and gives a set of clusters to the classifier network with multiple heads for the multiple predictions. The classifier network then gives its confidence back to OHSLIC and also outputs the labels for the different clusters. The number of clusters used in OHSLIC can be controlled by the clusters controller. | 2501.10199v1 | 4 |
|
Figure 5:Illustrative figure of the Classifier model. The model is made of a 1D CNN feature extractor and 4 heads, which are MLPs of different sizes. The output of the model is a 1D array of length 3 for the three different parameters and the confidence of the segmenation. | 2501.10199v1 | 5 |
|
Figure 6:Inference time for processing a line in ms with respect to the number of clusters for OHSLIC-C with and without confidence. These times come from running the algorithm on a NVIDIA Jetson Nano. | 2501.10199v1 | 6 |
|
Figure 7:Average regression performance with respect to the number of clusters from OHSLIC and the Dice score. | 2501.10199v1 | 7 |
|
Figure 8:Example of the OHSLIC-C-C segmentation, with the top panel being a fake RGB image of the HSI. With the segmentation panels being chlorophyll, carotenoid, and anthocyanin respectively. | 2501.10199v1 | 8 |
|
Figure 10:Example of how confidence changes the output of the model. On the left panel the fake RGB image of the dataset is visible. The middle panel is the normal segmentation without confidence and the right panel is the predictions with confidence. | 2501.10199v1 | 10 |
|
Figure 1:The network architectures adopted in existing methods and the associated challenges.red(a) Visualization of the network architectures for the methods described in (b). Detailed structural information is highlighted in the blue and orange boxes, representing the CNN-based bottleneck (GlobalG), Transformer-based bottleneck (SwinG), and the proposed SEC-CES-bottleneck. (b) Image-to-Image translation network based on Encoder-Bottleneck-Decoder paradigm. The proposed SCB has a similar hierarchical structure to GlobalG and SwinG but is not a simple stacking of RMs and SMs. It consists of SEC and CES cross-combination, considering both CNN and Transformer. | 2501.10197v1 | 1 |
|
Figure 5:Visualization results of the feature maps before and after the action
of IGC.xisubscript𝑥𝑖x_{i}italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTandxCESsubscript𝑥𝐶𝐸𝑆x_{CES}italic_x start_POSTSUBSCRIPT italic_C italic_E italic_S end_POSTSUBSCRIPTare the results before the optimization of parametersa𝑎aitalic_aandb𝑏bitalic_b.xi′superscriptsubscript𝑥𝑖′x_{i}^{\prime}italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPTandxCES′superscriptsubscript𝑥𝐶𝐸𝑆′x_{CES}^{\prime}italic_x start_POSTSUBSCRIPT italic_C italic_E italic_S end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPTare the feature maps after IGC. | 2501.10197v1 | 5 |
|
Figure 6:Visualization results of different methods for information asymmetric image translation on SEN12 dataset (scene-level). | 2501.10197v1 | 6 |
|
Figure 7:Visualization results of different models for information asymmetric image translation on Sketch2Anime dataset (instance-level). | 2501.10197v1 | 7 |
Subsets and Splits