image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Givenin-the-wildcaptures (a), our model enables transient objects removal (b) and scene reconstruction with variable appearances (c). | 2403.10427v2 | 1 |
|
Figure 3:Variance histogram analysis of the Gaussians’ opacity variationΔα~𝐈Δsuperscript~𝛼𝐈\Delta\mathbf{\tilde{\alpha}^{I}}roman_Δ over~ start_ARG italic_α end_ARG start_POSTSUPERSCRIPT bold_I end_POSTSUPERSCRIPTw.r.t training images for a dynamic (i.e. containing transient objects) scene (top, Trevi Fountain from Phototourism[6]); and a static scene (bottom, Locomotive from Tanks&Temples[9]).
The left column views are rendered using only static Gaussians (i.e having𝐕𝐚𝐫[Δα~𝐈]=0𝐕𝐚𝐫delimited-[]Δsuperscript~𝛼𝐈0\mathbf{Var}\left[\Delta\mathbf{\tilde{\alpha}^{I}}\right]=0bold_Var [ roman_Δ over~ start_ARG italic_α end_ARG start_POSTSUPERSCRIPT bold_I end_POSTSUPERSCRIPT ] = 0) whereas right column views are rendered using all Gaussians. | 2403.10427v2 | 3 |
|
Figure 6:Appearance transfer –SWAG is able to render a scene at any viewpoint with the visual appearance of any training image thanks to the learned appearance embeddings. | 2403.10427v2 | 6 |
|
Figure 7:Appearance interpolation –Interpolations between the appearance embeddings of two training images (left, right). | 2403.10427v2 | 7 |
|
Figure 8:Ablation study –Visual comparison of 3DGS, SWAG, and the two variants used in our ablation study: SWAG-T (no appearance) and SWAG-A (no transient). | 2403.10427v2 | 8 |
|
Figure 9:(a)Ground-Truth(b)3DGS[8](c)Affine color transformation prediction(d)Positional Encoding of 3D Gaussian centers(e)SWAG | 2403.10427v2 | 9 |
|
Figure 1:End point error (EPE) v.s. frame per second (FPS) throughput on a common computing platform (Nvidia RTX 2080). Individual points represents a broad class of optical flow methods. Our algorithm is comparable in accuracy but significantly better (close to an order of magnitude) in terms of its computational complexity. All models trained solely on FlyingThings and FlyingChairs. | 2403.10425v1 | 1 |
|
Figure 2:Optical flow results of NeuFlow: on the left is a result from the standard KITTI dataset. On the right are results from a UAS flight overlow-contrast glacier images in the Arctic. Our approach is notable for both computational efficiency and speed as well as accuracy, as shown in Fig.1. | 2403.10425v1 | 2 |
|
Figure 3:NeuFlow Architecture: We begins with a shallow CNN backbone. The backbone outputs feature vectors at 1/8 and 1/16 scale for both images. The feature vectors at 1/16 scale are then fed into two cross-attention layers for global matching. The resulting flow is passed into a self-attention layer for flow propagation based on feature self-similarity. Subsequently, the flow is upsampled to obtain 1/8 resolution flow. We wrap the 1/8 features with the flow and perform local refinement within a 7x7 window. The refined 1/8 flow is then upsampled to obtain full-resolution flow using a convex upsampling module, which additionally requires 1/8 features from image one. | 2403.10425v1 | 3 |
|
Figure 4:NeuFlow Shallow CNN Backbone: Initially, we downsample the image into different scales, ranging from 1/1 scale to 1/16 scale. Subsequently, we extract features using a CNN block. The feature vectors at 1/1, 1/2, 1/4, and 1/8 scales are concatenated into a single 1/8 feature vector. Then, another CNN block is employed to merge the 1/8 feature vector with the 1/16 feature vector, resulting in a 1/16 feature vector. The 1/16 feature vector is utilized for global attention, while the 1/8 feature vector is employed for local refinement. The CNN block consists solely of two CNN layers along with activation functions and normalization layers. The kernel size and stride of the CNN layers depend on the input and output dimensions upstream and downstream of the network. An additional 1/8 feature is extracted from the full-resolution image to perform convex upsampling. | 2403.10425v1 | 4 |
|
Figure 5:End point error (EPE) v.s. frame per second (FPS) on Nvidia RTX 2080 while outputting 1/8 resolution flow. All models trained solely on FlyingThings and FlyingChairs. NeuFlow is optimized for accuracy and efficiency at 1/8 resolution, thus we gain more advantage compared to full resolution flow. | 2403.10425v1 | 5 |
|
Figure 1:The proposed quantization scheme with quantization intervalℓℓ\ellroman_ℓ. The star represents a value to be quantized, and it is located in the quantization interval of[0,ℓ]0ℓ[0,\ell][ 0 , roman_ℓ ]under level-set 1 and[0.5ℓ,1.5ℓ]0.5ℓ1.5ℓ[0.5\ell,1.5\ell][ 0.5 roman_ℓ , 1.5 roman_ℓ ]under level-set 2.
At any even-number iteration (k𝑘kitalic_kis even), the star value will be quantized to either 0 orℓℓ\ellroman_ℓ, with respective probabilities provided in [7]. At any odd-number iteration (k𝑘kitalic_kis odd), the star value will be quantized to either0.5ℓ0.5ℓ0.5\ell0.5 roman_ℓor1.5ℓ1.5ℓ1.5\ell1.5 roman_ℓ, with respective probabilities given in [8]. | 2403.10423v1 | 1 |
|
Figure 4:Trajectories of all five agents when initialized on the saddle point (0,0). Note that all trajectories overlap with each other, implying perfect consensus among the agents. | 2403.10423v1 | 4 |
|
Figure 5:Comparison of the objective function value between the proposed Algorithm1and the existing algorithm DGD in(50). | 2403.10423v1 | 5 |
|
Figure 6:Comparison of the distance between learned parameters and the actual optimal solutionx∗superscript𝑥∗x^{\ast}italic_x start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT(obtained using centralized optimization). The learned parameters in our algorithm are represented asxAlg1subscript𝑥𝐴𝑙𝑔1x_{Alg1}italic_x start_POSTSUBSCRIPT italic_A italic_l italic_g 1 end_POSTSUBSCRIPT, and the learned parameters in the existing algorithm DGD are represented byxDGDsubscript𝑥𝐷𝐺𝐷x_{DGD}italic_x start_POSTSUBSCRIPT italic_D italic_G italic_D end_POSTSUBSCRIPT. It can be seen that our algorithm does converge to a better solution than DGD. | 2403.10423v1 | 6 |
|
Figure 7:Comparison of training accuracy between the proposed algorithm and a commonly used algorithm CDSGD from(52). | 2403.10423v1 | 7 |
|
Figure 8:Comparison of loss function value between the proposed algorithm and a commonly used algorithm CDSGD from(52). | 2403.10423v1 | 8 |
|
Figure 9:Comparison of reconstruction error in tensor decomposition between the proposed Algorithm 1 and the existing algorithm DGD in(50). | 2403.10423v1 | 9 |
|
Figure 10:Comparison of the reconstruction error in Robust PCA between the proposed Algorithm 1 and the existing algorithm Fast RPCA in(57). | 2403.10423v1 | 10 |
|
Figure 1.Taxonomy of Explainable AI according to(Guidotti et al.,2018b). In this research, we focus on gradient based explanations in feature attribution. | 2403.10415v1 | 1 |
|
Figure 3.The differences amongbackpropagation,deconvolutional network,guided backpropagation, andRectGradlie in the implementation of ReLU in the backward pass. | 2403.10415v1 | 3 |
|
Figure 4.The chronological evolution of gradients based explanations is depicted in the publication timeline. | 2403.10415v1 | 4 |
|
Figure 1:An illustration of our HyCTAS search space. It preserves the multi-resolution and multi-branch proprieties of HRNet by introducing two types of searchable components: cells and nodes, which extend HRNet with multi-module fusion ability. | 2403.10413v1 | 1 |
|
Figure 2:An illustration of our searchable modules. We design a memory-efficient self-attention module and a light-weight convolution module and search for the best combination. | 2403.10413v1 | 2 |
|
Figure 3:An illustration of our searched architectures with multiple branches. The red dot indicates the location of the self attention module. | 2403.10413v1 | 3 |
|
Figure 1.Centralised architecture (left) makes it difficult for users to have control over their data and use their data across applications. Decentralised architecture (right) decouples data from applications, giving users control over their data and making it easier for them to use their data across apps. | 2403.10408v1 | 1 |
|
Figure 2.Example Web application data flow in SocialGenPod. Alice allows Bob to interact with or query her personal data through a Web app. The app uses a retrieval service to find permitted documents in Alice’s Solid Pod that are most relevant to Bob’s query. The app then uses one of the available generative AI models to generate a contextualised response for Bob. | 2403.10408v1 | 2 |
|
Figure 3.RAG with private data in SocialGenPod. A chatbot without RAG (top) is unable to answer a user’s query. The same chatbot with RAG (bottom) produces a better response. | 2403.10408v1 | 3 |
|
Figure 1:We present a goal-conditioned imitation learning framework for sculpting clay that uses point-cloud state representation. We find that our system is much faster at test time compared to traditional methods planning with a dynamics model. However, our system is more limited in scope as imitation learning does not generalize well to unseen goals. | 2403.10401v1 | 1 |
|
Figure 2:The pipeline of SculptDiff. The state and goal point clouds are encoded with PointBERT[18]and a linear projection head to create a latent conditioning observation along with the previous action executed by the robot. The latent state and goal observations as well as the previous action are the conditioning information used to condition the denoising diffusion process for diffusion policy[17]to generate the predicted action sequence. | 2403.10401v1 | 2 |
|
Figure 3:The experimental setup includes 4 Intel RealSense D415 RGB-D cameras mounted to a camera cage to reconstruct the clay point cloud. An additional Intel RealSense D455 overhead camera is used to record experimental videos. We fit the robot with 3D printed fingertips and an elevated stage similar to those in[4]. We assume the clay always remains centered and fixed to the elevated stage throughout the experiments. | 2403.10401v1 | 3 |
|
Figure 4:The TSNE embeddings for PointBERT with different training strategies on the X demonstration dataset. The colorbar of state index indicates the state number ranging from 0 to 7 for the states in each demonstration trajectory for the X shape. a) PointBERT pre-trained on ShapeNet, b) PointBERT finetuned with diffusion policy, c) PointBERT finetuned with ACT policy, and d) PointBERT finetuned with VINN policy. | 2403.10401v1 | 4 |
|
Figure 5:The final shapes created by the policies trained with point cloud inputs for a single shape goal. For the target point cloud (representing the shapes created by the human oracle using hands) on the left-most column, the lightness of each point is correlated with the point’s z-value to visualize depth. While both human oracles create the best shapes, point cloud diffusion policy is able to successfully create the closest matches to the human demonstrations. | 2403.10401v1 | 5 |
|
Figure 2:The pipeline of Isotropic3D. Neural Radiance Field (NeRF) utilizes volume rendering to extract four orthogonal views, which are subsequently augmented with random Gaussian noise. These views, along with noise-free reference images, are then transferred to a multi-view diffusion model for predicting added noise. Note that, weset the timestept𝑡titalic_tto zeroat the corresponding position of noise-free reference images. The framework that generates consistent multi-view images from only a single CLIP embedding can be aligned with the input view while retaining the consistency of the output target view. Finally, NeRF yields high-quality 3D content optimized by rendered images via Score Distillation Sampling (SDS).ℒ𝒮𝒟𝒮subscriptℒ𝒮𝒟𝒮\mathcal{L_{SDS}}caligraphic_L start_POSTSUBSCRIPT caligraphic_S caligraphic_D caligraphic_S end_POSTSUBSCRIPTcan refer toEq.7. | 2403.10395v1 | 2 |
|
Figure 3:View-Conditioned Multi-view Diffusion pipeline. Our training process is divided into two stages. In the first stage (Stage1), we fine-tune a text-to-3D diffusion model by substituting its text encoder with an image encoder, by which the model preliminarily acquires image-to-image capabilities. Stage1-a and Stage1-b are the single-view diffusion branch and the multi-view diffusion branch for the first stage respectively. In the second stage (Stage2), we perform fine-tuning multi-view diffusion model integrated Explicit Multi-view Attention (EMA). EMA combines noisy multi-view images with the noise-free reference image as an explicit condition. Stage2-a and Stage2-b are diffusion branches for the second stage.
During inference, we only need to send theCLIP embeddingof the reference image and camera pose to generate consistent high-quality images from multiple perspectives. | 2403.10395v1 | 3 |
|
Figure 4:Illustration of the Explicit Multi-view Attention (EMA). “View-Input” is a feature map of the noise-free reference image. “View 1” and “View 1∼similar-to\sim∼4” are feature maps of noisy rendered views. “Alternative” means a 30% chance of using single-view diffusion (Stage2-a) and a 70% chance of training with the multi-view diffusion branch (Stage2-b). | 2403.10395v1 | 4 |
|
Figure 5:Qualitative comparison of synthesizing novel views with baseline models[21,22]on GSO[6]and randomly collected images. | 2403.10395v1 | 5 |
|
Figure 6:Qualitative comparisons of 3D Generation with baseline models. We conducted verification on GSO[6]and randomly collected images.Isotropic3Dis capable of generating more regular geometry, detailed texture, and less flat compared with Zero123 and Magic123. A video of this result is available athttps://isotropic3d.github.io/. | 2403.10395v1 | 6 |
|
Figure 8:Qualitative comparisons in different settings.CCRis denote aschannel-concatenatereference image.NOTHINGmeans that it does not generate anything. A video of this result is available athttps://isotropic3d.github.io/. | 2403.10395v1 | 8 |
|
Figure 9:Addtional Results for qualitative comparisons of Isotropic3D with baseline models. A video of this result is available on our project website. | 2403.10395v1 | 9 |
|
Figure 3:(a) and (b) present the class probabilities predicted on a white image using the proposed algorithm. (c) and (d) present the confusion matrices of the class predictions on test samples. | 2403.10391v2 | 3 |
|
Figure 7:Confusion matrices of pseudo-labels generated by (a) FixMatch, (b) FixMatch+CDMAD, (c) ReMixMatch, and (d) ReMixMatch+CDMAD trained on CIFAR-10-LT underγl=100subscript𝛾𝑙100\gamma_{l}=100italic_γ start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT = 100andγu=100subscript𝛾𝑢100\gamma_{u}=100italic_γ start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPT = 100. | 2403.10391v2 | 7 |
|
Figure 8:Confusion matrices of the class predictions on the test set of CIFAR-10 using (a) FixMatch, (b) FixMatch+CDMAD, (c) ReMixMatch, and (d) ReMixMatch+CDMAD trained on CIFAR-10-LT underγl=100subscript𝛾𝑙100\gamma_{l}=100italic_γ start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT = 100andγu=100subscript𝛾𝑢100\gamma_{u}=100italic_γ start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPT = 100. | 2403.10391v2 | 8 |
|
Figure 9:Confusion matrices of pseudo-labels generated by (a) FixMatch, (b) FixMatch+CDMAD, (c) ReMixMatch, and (d) ReMixMatch+CDMAD trained on CIFAR-10-LT underγl=100subscript𝛾𝑙100\gamma_{l}=100italic_γ start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT = 100andγu=1subscript𝛾𝑢1\gamma_{u}=1italic_γ start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPT = 1. | 2403.10391v2 | 9 |
|
Figure 10:t-SNE of representations obtained for the test set of CIFAR-10 using (a) FixMatch, (b) FixMatch+CDMAD, (c) ReMixMatch, and (d) ReMixMatch+CDMAD trained on CIFAR-10-LT underγl=100subscript𝛾𝑙100\gamma_{l}=100italic_γ start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT = 100andγu=1subscript𝛾𝑢1\gamma_{u}=1italic_γ start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPT = 1. | 2403.10391v2 | 10 |
|
Figure 11:Confusion matrices of the class predictions on the test set of CIFAR-10 using (a) FixMatch and (b) FixMatch+CDMAD trained on CIFAR-10-LT underγl=100subscript𝛾𝑙100\gamma_{l}=100italic_γ start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT = 100andγu=1subscript𝛾𝑢1\gamma_{u}=1italic_γ start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPT = 1. | 2403.10391v2 | 11 |
|
Figure 1:One kernel atxssubscript𝑥𝑠x_{s}italic_x start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT(dotted kernel) or two kernels atxisubscript𝑥𝑖x_{i}italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTandxjsubscript𝑥𝑗x_{j}italic_x start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT(left and right) lead to the same summed estimate atxssubscript𝑥𝑠x_{s}italic_x start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT.
This shows a figure consisting of different types of lines.
Elements of the figure described in the caption should be set in italics, in parentheses, as shown in this sample caption.
The last sentence of a figure caption should generally end with a full stop, except when the caption is not a full sentence. | 2403.10390v2 | 1 |
|
Figure 1:What is the Optimal Target for rPPG?Most rPPG models are trained with PPG targets from the finger tip; however, waveform morphology differs in phase and morphology at different sites across the body. We show that using finger tip PPG targets is less optimal than using PPG measured from the face. | 2403.10582v1 | 1 |
|
Figure 2:Experimental Apparatus.We used a customarily designed data collection apparatus that simultaneously collects contact reflectance PPG measurements from the face (forehead) (a) and finger (b) synchronized with video recordings of the face. | 2403.10582v1 | 2 |
|
Figure 3:Test Loss on Waveform Predictions.Mean squared error (MSE) between the predicted PPG waveforms and reference contact sensor measurements. Error bars show the standard deviation across subjects. MSE Lower = Better | 2403.10582v1 | 3 |
|
Figure 4:Waveform Examples from Two Participants.Contact sensor (gray) and rPPG predictions (red) for test data from two participants using the DeepPhys model (top) trained on finger and tested on face, (second top) trained on face and tested on finger, (third top) trained on finger and tested on face, (fourth top) trained on face and tested on face. | 2403.10582v1 | 4 |
|
Figure 1:Our proposed SSL framework. The right panel shows 8 popular SSL algorithms, each of which has seemingly different model graphs, network architectures, losses, and augmentation strategies. We show that all of these methods are instances of a single unifed framework (left), each having different architecture hyperparameters for each module, augmentation hyperparameters, or losses. | 2403.05726v1 | 1 |
|
Figure 1:Color-coded p-values from backtestingtsGTonelectricity. Each rectangle represents the results for backtestingtsGTon one of the following levels:50%percent5050\%50 %,75%percent7575\%75 %, and95%percent9595\%95 %. The height of a rectangle equalsH=24𝐻24H=24italic_H = 24, corresponding to the number of prediction steps, and a width equals toS=321𝑆321S=321italic_S = 321, the number of time series inelectricity. | 2403.05713v3 | 1 |
|
Figure 2:Visualization of data from each of the four real-life datasets. The first row displays a year’s worth of data for a selected series. The second and the third rows zoom in on the suffix of those trajectories. | 2403.05713v3 | 2 |
|
Figure 3:Reconstruction of apple sketches bytsGT. The ground truth images are shown in the top row. In the bottom row, the partial ground truth trajectory used to prompttsGTis shown in blue, and the reconstructed trajectory is highlighted in orange. | 2403.05713v3 | 3 |
|
Figure 4:Reconstruction of banana sketches bytsGT. The ground truth images are shown in the top row. In the bottom row, the partial ground truth trajectory used to prompttsGTis shown in blue, and the reconstructed trajectory is highlighted in orange. | 2403.05713v3 | 4 |
|
Figure 5:Reconstruction of castle sketches bytsGT. The ground truth images are shown in the top row. In the bottom row, the partial ground truth trajectory used to prompttsGTis shown in blue, and the reconstructed trajectory is highlighted in orange. | 2403.05713v3 | 5 |
|
Figure 6:Reconstruction of flower sketches bytsGT. The ground truth images are shown in the top row. In the bottom row, the partial ground truth trajectory used to prompttsGTis shown in blue, and the reconstructed trajectory is highlighted in orange. | 2403.05713v3 | 6 |
|
Figure 3:A comparison of different feature attentive mechanisms. The proposed Shift-Parser
(c) takes advantages of both spatial domain methods (a) and frequency domain method (b), which groups the features in frequency domain while also maintaining the spatial constraint. | 2403.05703v1 | 3 |
|
Figure 4:The visualized results of Shift-Parser mask in each stage (mapping back to the raw image). The focal parts are kept and the context parts are masked. For the small focal parts, we zoom in the region on the up-left/right corners for clarity. From top to down are wildlife/vehicles, medical image, and artwork classification respectively. | 2403.05703v1 | 4 |
|
Figure 5:Pro-NeXt Gflops are strongly correlated with performance. We plot the Gflops of each of our Pro-NeXt models and each model’s mean performance after 200K training steps. | 2403.05703v1 | 5 |
|
Figure 1:Comparison of ourT2AVwith state-of-the-art methodson the proposedT2AV-Benchin terms of FAD, FAVD, FATD, and FA(VT)D for video-aligned text-to-audio generation.
Our method significantly outperforms previous baselines in terms of all metrics (lower is better). | 2403.07938v1 | 1 |
|
Figure 2:Illustration of the proposed framework for Text-to-audio generation aligned with videos (T2AV).The Audio-Visual ControlNet aggregates temporal video features as the condition in the text-based latent diffusion models.
Then, a contrastive language-audio pretraining objective across each temporal location is applied to match visual-aligned text embeddings with audio features.
After visual-aligned CLAP pre-training, we directly extract text embeddings with Audio-Visual ControlNet as the condition for latent diffusion models to achieve Text-to-Audio generation with visual alignment and temporal consistency. | 2403.07938v1 | 2 |
|
Figure 3:Qualitative comparisons with AudioLDM on video-aligned TTA generation.The proposed T2AV produces more accurate and aligned audio for target videos. | 2403.07938v1 | 3 |
|
Figure 1:The schematic of a proposed framework for detecting glaucoma. The framework uses a large Vision Transformer (ViT-large) feature extractor called RETFound, which was pre-trained on a large OCT dataset. The symbol⊕direct-sum\oplus⊕represents the concatenation of spatial states at timestepi𝑖iitalic_i, combining information from anterior and posterior spatial locations to capture long-range spatial dependencies. The framework includes a Bidirectional Gated Recurrent Unit (GRU), Adaptive Max Pooling (AMP), and a Fully Connected (FC) layer. | 2403.05702v1 | 1 |
|
Figure 1:Shortened examples of the LLM evaluation tasks. Correlations are based on answers across multiple stimuli. | 2403.05701v2 | 1 |
|
Figure 4:Spearman correlation between model answers and human answers forExperiment 2.** forp<0.05𝑝0.05p<0.05italic_p < 0.05,bold= highest correlation, N/A = model always returns the same score | 2403.05701v2 | 10 |
|
Figure 1:Examples fromSeeGULL Multilingual. Lang. (Language): es: Spanish, fr: French, it: Italian, sw: Swahili, fr: French; S: # of annotators (out of 3) who reported it as a stereotype; O: mean offensiveness rating of the stereotype. | 2403.05696v1 | 1 |
|
Figure 2:Example differences in known stereotypes in the same language across two different countries. S(Cisubscript𝐶𝑖C_{i}italic_C start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT) is the # annotators marking the tuple as stereotype in countryCisubscript𝐶𝑖C_{i}italic_C start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT. Countries are denoted by their ISO codes. | 2403.05696v1 | 2 |
|
Figure 3:Offensive Annotations for nationalities of the world. We take all the stereotypes along the nationality axis, and find the averagemean offensive score, corresponding to each country. The countries having the darker shades of red, have on an average, more offensive stereotypes associated with them. | 2403.05696v1 | 3 |
|
Figure 4:Example of highly offensive stereotypes. The columncountrydenotes the country of annotation. | 2403.05696v1 | 4 |
|
Figure 5:Example of evaluation prompt in Bengali and English translation. The stereotypical identity associated with the blue attribute is highlighted in orange. | 2403.05696v1 | 5 |
|
Figure 6:Endorsement of stereotypes varies by language and place. Endorsements per language and country are aggregated across all models. | 2403.05696v1 | 6 |
|
Figure 1:Electroluminiscence image of a single mono-Si solar cell containing type A, B, and C cracks. | 2403.05694v1 | 1 |
|
Figure 6:TensorFlow-Lite (TF-lite) and full integer quantization performance (FIQ) analysis in environment B. | 2403.05694v1 | 14 |
|
Figure 2:Action history and reaction wheel speeds when deploying under policyπ0subscript𝜋0\pi_{0}italic_π start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT(top) andπ1subscript𝜋1\pi_{1}italic_π start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT(bottom) from a fixed initial condition. The red highlight shows when the spacecraft has access to the target, the blue highlights show when the spacecraft is in Momentum Dumping (RW Desat) Mode. Note that policyπ1subscript𝜋1\pi_{1}italic_π start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT(trained inφ0L∧φSsubscript𝜑0𝐿subscript𝜑𝑆\varphi_{0L}\land\varphi_{S}italic_φ start_POSTSUBSCRIPT 0 italic_L end_POSTSUBSCRIPT ∧ italic_φ start_POSTSUBSCRIPT italic_S end_POSTSUBSCRIPT) keeps the spacecraft safe after imaging the target whereas policyπ0subscript𝜋0\pi_{0}italic_π start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT(trained on onlyφ0Lsubscript𝜑0𝐿\varphi_{0L}italic_φ start_POSTSUBSCRIPT 0 italic_L end_POSTSUBSCRIPT) prioritizes imaging over spacecraft survival. | 2403.05693v3 | 2 |
|
Figure 1:Overview of our proposedUEMsemantic feature learning framework forU22{}^{2}start_FLOATSUPERSCRIPT 2 end_FLOATSUPERSCRIPTCDR. In the first stage of Intra-Domain Semantic-Enhanced Learning,UEMestablishes a unified prototypical structure across domains, which is driven and enhanced byIPMandSEL, respectively. Then, in the second stage of Cross-Domain Semantic-Matched Learning,SPDAis used to align domains while preserving the built prototypical structure, andSN22{}^{2}start_FLOATSUPERSCRIPT 2 end_FLOATSUPERSCRIPTMcan achieve more accurate cross-domain categorical pairing. | 2403.05690v1 | 1 |
|
Figure 4:Retrieval results of our framework on Office-31 (A→→\rightarrow→D) and Office-Home (A→→\rightarrow→C) in Closet Unsupervised Cross-Domain Retrieval. | 2403.05690v1 | 3 |
|
Figure 3:The overall architecture of proposedSGRRG𝑆𝐺𝑅𝑅𝐺SGRRGitalic_S italic_G italic_R italic_R italic_G. The basic visual features obtained from the visual feature extractor are fed into the vision encoder, scene graph construction module and the scene graph encoder to obtain the encoded global and region level representations respectively. These representations are then taken as the input source for the proposed scene graph-aided decoder to predict the reports. | 2403.05687v1 | 2 |
|
Figure 4:An example of reports generated by different models. Most medical terms are highlighted. Abnormal terms are in bold. | 2403.05687v1 | 3 |
|
Figure 1:A stylized depiction of Alignment Studio with its three components: Framers, Instructors, and Auditors. | 2403.09704v1 | 1 |
|
Figure 2:A realization of the Alignment Studio software architecture, starting with policy documents. End-to-end software testing and documentation is recommended, but
implementations need not include all components. | 2403.09704v1 | 2 |
|
Figure 3:Creating instruction style seed data from policy documents (left) and using it to generate synthetic data using LLMs in a few-shot setting (right). | 2403.09704v1 | 3 |
|
Figure 1:“GPTRadScore” framework for the auto-evaluation of LLM descriptions of CT-based findings. CT slices with outlined lesions were fed to vision-based LLMs that generated a description of the finding. They were then evaluated against the gold-standard sentences by a clinician, with NLG metrics, and auto-evaluation with GPT-4. | 2403.05680v2 | 1 |
|
Figure 2:Comparison of the responses from multi-modal LLMs (using CoT reasoning) for a renal cyst in the right kidney.Red,blueandpurplefonts denote incorrect, correct, and uncertain descriptions respectively. | 2403.05680v2 | 2 |
|
Figure 3:Heatmap of pairwise Pearson’s Correlation Coefficient among various grading scores; traditional metrics, Clinician evaluations and GPTRadScore for Gemini Pro Vision, GPT-4V, LLaVA-Med, RadFM, and RadFM (FT). Color intensity indicates the strength of correlation, with darker shades representing higher correlation. | 2403.05680v2 | 3 |
|
Figure 4:Comparison of results of abnormality characterization by GPT-4V, Gemini Pro Vision, LLaVA-Med, RadFM, and RadFM (FT) with bounding boxes (bbox) vs. without bounding boxes (w/o bbox). Color mapping = {orange: ‘Incorrect’, beige: ‘Partially Correct’, teal: ‘Correct’, white: ‘Not Applicable’}.x𝑥xitalic_x-axis denotes scores{x∈ℝ|0<x<1}conditional-set𝑥ℝ0𝑥1\{\ x\in\mathbb{R}\ |\ 0<x<1\ \}{ italic_x ∈ blackboard_R | 0 < italic_x < 1 },N𝑁Nitalic_N= 500 samples. | 2403.05680v2 | 4 |
|
Figure 1:Based on three performance-centric observations (O1∼similar-to\sim∼O3), PipeRAG combines a system-aware algorithm integrating pipeline parallelism (S1) with flexible retrieval intervals (S2) and an algorithm-aware retrieval system guided by a performance model (S3). | 2403.05676v1 | 1 |
|
Figure 4:The effect of database sizes and retrieval strategies on language modeling perplexity (lower perplexity means higher quality). | 2403.05676v1 | 4 |
|
Figure 5:Perplexity of retrieval-augmented generation when applying various retrieval intervals and search space configurations (nprobe𝑛𝑝𝑟𝑜𝑏𝑒nprobeitalic_n italic_p italic_r italic_o italic_b italic_e). | 2403.05676v1 | 8 |
|
Figure 6:PipeRAG significantly outperformsRetroon the latency-perplexity Pareto frontier (lower latency and perplexity are better). | 2403.05676v1 | 12 |
|
Figure 7:Trends in PipeRAG efficiency when deployed on future hardware that enables faster retrieval or inference. | 2403.05676v1 | 16 |
|
Figure 8:Even if the baseline model supports flexible retrieval intervals (Retro+), PipeRAG still significantly outperforms it in efficiency thanks to the proposed pipeline parallelism. | 2403.05676v1 | 17 |
|
Figure 10:The latency-perplexity comparison between PipeRAG versusRetrogiven the same retrieval intervals. | 2403.05676v1 | 22 |
|
Figure 1:Illustrative occlusion scenarios from the Boreas dataset[burnett_ijrr23]where the ego vehicle encounters occlusion caused by a truck. Camera images for two scenarios are shown on the left, with corresponding lidar scans captured at the same time on the right. Scenario a) precedes b) by 0.3 seconds. The ground planes in the lidar scans are removed for clarity. The pose of the ego vehicle and the region occluded by the truck are annotated with a red arrow and a gray sector, respectively. When the truck occludes the green-circled landmarks in b), theICPalgorithm may mistake the truck for the circled landmarks in its calculations, resulting in a significant error in the pose estimate. | 2403.05666v2 | 1 |
|
Figure 2:Overview of the proposed attack pipeline. Given an input point cloud𝑿𝑿\bm{X}bold_italic_X, the encoder extracts hierarchical features from it using set-abstraction modules from the PointNet++[qi2017pointnet]architecture. The decoder interpolates the extracted features and uses them along with the original point cloud to produce an adversarial point cloud𝑿advsubscript𝑿adv\bm{X}_{\text{adv}}bold_italic_X start_POSTSUBSCRIPT adv end_POSTSUBSCRIPTto attackICP. The encoder and decoder together form the generator𝑮𝑮\bm{G}bold_italic_G, which is trained using reconstruction loss and adversarial loss. | 2403.05666v2 | 4 |
|
Figure 3:A qualitative comparison between the original point clouds and the adversarial point clouds produced by our model and the baselines. Columns 2 through 6 showcase the adversarial point clouds resulting from applying our models and the baselines to the original point clouds in column 1.
The adversarial point clouds are coloured according to the angle of the displacement vectors,𝑿adv−𝑿subscript𝑿adv𝑿\bm{X}_{\text{adv}}-\bm{X}bold_italic_X start_POSTSUBSCRIPT adv end_POSTSUBSCRIPT - bold_italic_X.
The heading pose error and the norm of the lateral and longitudinal pose errors ofICPinduced by each point cloud are indicated at the upper and lower right corners, respectively.The lateral and longitudinal pose errors are unitless as the point clouds are normalized. MinorICPpose errors are observed even when localizing the uncorrupted point clouds in column 1. This occurs because noise is added to the scans to prevent perfect alignment between the scans and the maps. | 2403.05666v2 | 5 |
|
Figure 4:Worst-case translation pose errors over a route estimated by our model when allowed up to1mtimes1m1\text{\,}\mathrm{m}start_ARG 1 end_ARG start_ARG times end_ARG start_ARG roman_m end_ARGof perturbation.Locations are coloured based on the magnitude of their worst-case translation pose error.Errors are capped at3mtimes3m3\text{\,}\mathrm{m}start_ARG 3 end_ARG start_ARG times end_ARG start_ARG roman_m end_ARGfor better visualization.Some locations, such asa,b,d,𝑎𝑏𝑑a,b,d,italic_a , italic_b , italic_d ,ande𝑒eitalic_e, are much more prone to attacks than others such asc𝑐citalic_c. | 2403.05666v2 | 6 |
|
Figure 5:ICPerrors induced by our method minus those estimated by[laconte2023certifying]on the same trajectory. A positive difference means ours finds a more detrimental perturbation pattern than[laconte2023certifying].λ=1m𝜆times1m\lambda=$1\text{\,}\mathrm{m}$italic_λ = start_ARG 1 end_ARG start_ARG times end_ARG start_ARG roman_m end_ARGfor both approaches. Left: the trajectory coloured by the error difference. Right: a violin plot of the distribution of differences, with the interquartile range denoted in white. | 2403.05666v2 | 7 |
|
Figure 6:An extended comparison between the original point clouds and the adversarial point clouds produced by our model and the baselines.Columns 2 through 6 showcase the adversarial point clouds resulting from applying our models and the baselines to the original point clouds in column 1.
The adversarial point clouds are coloured according to the angle of the displacement vectors,𝑿adv−𝑿subscript𝑿adv𝑿\bm{X}_{\text{adv}}-\bm{X}bold_italic_X start_POSTSUBSCRIPT adv end_POSTSUBSCRIPT - bold_italic_X.
The heading pose error and the norm of the lateral and longitudinal pose errors ofICPinduced by each point cloud are indicated at the upper and lower right corners, respectively.The lateral and longitudinal pose errors are unitless as the point clouds are normalized. MinorICPpose errors are observed even when localizing the uncorrupted point clouds in column 1. This occurs because noise is added to the scans to prevent perfect alignment between the scans and the maps. | 2403.05666v2 | 8 |
Subsets and Splits