image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:A rectangular prism typically offers greater stability compared to a cylinder or a sphere.
The ease with which a cylinder rolls is contingent upon its orientation, while a sphere possesses the capability to roll in any direction. | 2403.10689v1 | 1 |
|
Figure 5:PCA on latent space after training the 1st phase vision module, which is transferred to the 2nd phase haptic-audio module.
The latent space represents the object shape, orientation and position. | 2403.10689v1 | 5 |
|
Figure 6:Assessment of prediction error throughout 20,000 epochs of training, contrasting the performance between the proposed method and the baseline approach. | 2403.10689v1 | 6 |
|
Figure 7:Position and orientation recognition result, showing the prediction error.
In a box plot, the cross shows the average, the the lower edge of the box represents the first quartile, the line drawn inside the box represents the second quartile, and the upper edge of the box represents the third quartile.
The vertical lines represent the maximum and minimum values and the dots are outliers. | 2403.10689v1 | 7 |
|
Figure 8:Object shape recognition results which show the number of correct and incorrect predictions. | 2403.10689v1 | 8 |
|
Figure 9:Tracking results observed over a 15-second duration (consisting of 150 time steps at a rate of 10 Hz) during the real robot experiment with the wood cylinder and employing the proposed model | 2403.10689v1 | 9 |
|
Figure 10:Snapshots of prediction results in the real robot experiment using untrained objects with the proposed model.
The arrows are showing the predicted shape with their colour, orientation with their direction, and position with the position they start. | 2403.10689v1 | 10 |
|
Figure 3:(1). Co-Segmenterincludes a transformer-like module and a mask decoder to produce the co-segmentation masks.(2). Obj-Segmenterconsists of the DINOv2 backbone, a transformer-like module, and a mask decoder to predict the object mask.(3). RA-Encodercontains the DINOv2 backbone, four3×3333\times 33 × 32D convolutional (Conv2D) layers with stride 2, a generalizable average pooling layer, and a fully connected (FC) layer. | 2403.10683v2 | 3 |
|
Figure 5:Qualitative evaluation on LINEMOD. We present the intermediate segmentation mask predictions (for localization) as well as the estimated 6D poses. Blue, green, and red boxes represent initial, refined, and ground truth poses, respectively. | 2403.10683v2 | 8 |
|
Figure 1:Left: Three examples of 2D ball detections in an accumulated event frame which serves as the input to theSpiking Neural Network(SNN) withground truth in greenand theestimated position in red.
Right: Five observed 2D trajectories in the camera frame of the event-based camera withground truth in greenand theestimated positions in red.
Background: The table tennis robot setup with the robot hitting back a table tennis ball in a rally. | 2403.10677v2 | 1 |
|
Figure 2:Our camera setup consisting offour frame-based cameras (in blue)andtwo event-based cameras (in red)with baselines of3333m to5555m.
Schematic is up to scale. | 2403.10677v2 | 2 |
|
Figure 3:The network has128128128128output neurons, which are split into two populations.
Each neuron in the first population represents anx𝑥xitalic_x-position, and each neuron in the second represents ay𝑦yitalic_y-position.
The target is also split into anx𝑥xitalic_x- andy𝑦yitalic_y-target, each setting one as the target activity for the correct neuron,0.50.50.50.5for the two adjacent neurons, and zero for all others.
Values are represented using brightness, with larger values being brighter. | 2403.10677v2 | 3 |
|
Figure 4:Left: Experiment setup for the offline experiment.
Right: Experiment setup for the online experiment. | 2403.10677v2 | 4 |
|
Figure 2:The architecture of the DLK-Net. DLK-Net consists of an encoder, a bottleneck, and a decoder. Two consecutive DLK blocks are used in each stage for feature extraction. Each DLK block consists of a DLK module and an MLP module. | 2403.10674v2 | 2 |
|
Figure 3:The architecture of the DLK-NETR. DLK-NETR utilized the same encoder as DLK-Net and D-Net. In this encoder, two consecutive DLK blocks are used in each stage for feature extraction. Each DLK block consists of a DLK module and an MLP module. This encoder is incorporated into a hybrid CNN-ViT architecture which is the same as UNETRHatamizadeh et al. (2022a), Swin UNETRHatamizadeh et al. (2021), UX NetLee et al. (2022), and VSmTransLiu et al. (2024). | 2403.10674v2 | 3 |
|
Figure 4:Qualitative comparison between D-Net and other methods ((c) VNet (d) nnU-Net (e) TransBTS (f) nnFormer (g) MedNext (f) SegFormer) across three public datasets, including the AMOS 2022 Multi-organ dataset, the MSD Brain Tumor dataset, and the MSD Hepatic Vessel Tumor dataset. | 2403.10674v2 | 4 |
|
Figure 5:Qualitative comparison between DLK-NETR and other methods which employ the same architecture ((c) UNETR (d) Swin UNETR (e) UX Net (f) VSmTrans) across three public datasets, including the AMOS 2022 Multi-organ dataset, the MSD Brain Tumor dataset, and the MSD Hepatic Vessel Tumor dataset. | 2403.10674v2 | 5 |
|
Figure 6:Qualitative comparison between Conv-ViT and DLK-ViT on AMOS multi-organ segmentation task. DLK-ViT shows the best segmentation quality than Conv-ViT, demonstrating the effectiveness and superior segmentation capabilities of DLK. | 2403.10674v2 | 6 |
|
Figure 7:(A) Qualitative comparison between nnU-net, nnU-Net incorporated with AFF (nnU-Net+AFF), and nnU-Net incorporated with DFF (nnU-Net+DFF) on AMOS multi-organ segmentation task. (B) Qualitative comparison between DLK-ViT, DLK-ViT incorporated with AFF (DLK-ViT+AFF), and DLK-ViT incorporated with DFF (DLK-ViT+DFF(DLK-Net)) on AMOS multi-organ segmentation task. The incorporation of DFF into nnU-Net and DLK-ViT improves their segmentation quality. Additionally, nnU-Net and DLK-ViT incorporated with DFF show better segmentation quality than those incorporated with AFF. | 2403.10674v2 | 7 |
|
Figure 8:Qualitative comparison of failure segmentation results from D-Net, DLK-NETR, and other SOAT methods across three public datasets. Two cases are shown from the AMOS multi-organ dataset and the Brain Tumor dataset, and one case is shown from the Hepatic Vessel Tumor dataset. | 2403.10674v2 | 8 |
|
Figure 1:Learned RFMP flows () from the base distribution () to the LASA datasets𝖲𝖲\mathsf{S}sansserif_Sand𝖶𝖶\mathsf{W}sansserif_W() on bothℝ2superscriptℝ2\mathbb{R}^{2}blackboard_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT(top) and𝒮2superscript𝒮2\mathcal{S}^{2}caligraphic_S start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT(bottom). The flow is conditioned on random observations𝒐𝒐\bm{o}bold_italic_ofrom the training dataset (). | 2403.10672v2 | 1 |
|
Figure 1:Summary of the regularization-based approach
to uncertainty quantification that underlies Hessian-free Laplace.Step 1:fit or obtain an existing fit of a predictive model;step 2:fit another model using a differently regularized objective (from scratch or warm-started);step 3:the difference in predictions of these two models is the predictive variance.
In Section3, we present various types of regularization that make this strategy successful. | 2403.10671v1 | 1 |
|
Figure 3:The uncertainty ranges of a random subset of 30 parameters of the neural network indicate that many parameters include 0 in their intervals. HFL recovers much of the structure of these uncertainties via additional regularization of the network parameters in the objective function. | 2403.10671v1 | 7 |
|
Figure 4:Absolute errors in predicted means for all the tasks with networks with added sparsity using the uncertainty quantification methods. | 2403.10671v1 | 8 |
|
Figure 2:a) Source training (Eq.1), b) Trigger set training (Eq.3), c) Feature regularization (Eq.8) d) Surrogate model (Eq.2). The proposed MAT identifies samples close to the decision boundary as the trigger set and adjusts the features of this set to align closely with the class center of the modified label. | 2403.10663v2 | 2 |
|
Figure 3:MAT significantly outperforms the margin-based approach in watermarking effectiveness when facing white-box attacks. | 2403.10663v2 | 3 |
|
Figure 4:A largeα𝛼\alphaitalic_αenhances feature regularization, thereby resulting in improved watermarking performance. | 2403.10663v2 | 4 |
|
Figure 4:A largeα𝛼\alphaitalic_αenhances feature regularization, thereby resulting in improved watermarking performance. | 2403.10663v2 | 5 |
|
Figure 5:MAT can achieve strong watermarking performance with a small trigger set consisting of multi-view data. | 2403.10663v2 | 6 |
|
Figure 1:Unified Vision: This figure provides a general overview of the seamless integration designed for joint semantic segmentation and depth estimation, ultimately contributing to the creation of a 3D scene. | 2403.10662v1 | 1 |
|
Figure 2:Network Architecture: Figure shows the detailed network architecture tailored for the precise prediction of both depth maps and segmentation maps. Notably, the network employs two critic discriminators during training, enhancing its ability to discern and refine the joint output representing depth and segmentation information. | 2403.10662v1 | 2 |
|
Figure 3:The first row shows a stereo pair from the Cityscapes dataset. The second row illustrates the dense disparity map (bottom right) and the dense depth map (bottom left), both generated using the advanced Cascaded Recurrent Stereo Matching Network (CREStereo)[32] | 2403.10662v1 | 3 |
|
Figure 4:Histogram of depth values of an instance in the Cityscapes dataset, visually highlighting the dominance of values within the0−100100-100 - 10meter range. The distribution’s peak and long tail motivate the logarithmic transformation. | 2403.10662v1 | 4 |
|
Figure 1:InterLUDE Framework. The labeled and unlabeled data are individually subjected to both weak and strong augmentations, followed by the backbone extracting feature embeddings. Embedding fusion then perturbs the embeddings as shown in Fig2.
The delta consistency loss is employed to regulate model behavior on labeled and unlabeled samples under the same augmentation changes. | 2403.10658v1 | 1 |
|
Figure 2:Illustration of Embedding Fusion with Batch Size of 2. This operation perturbe each embedding slightly by its immediate previous neighbor embedding. Showing the flattened embeddings for clear visualization. | 2403.10658v1 | 2 |
|
Figure 6:Effect of Different Batch Layout. Showing result on CIFAR-10. From left to right are CIFAR-10 with 40 labels, 250 labels and 4000 labels. | 2403.10658v1 | 5 |
|
Figure 2:Illustrations of LR importanceij,ntsuperscriptsubscript𝑖𝑗𝑛𝑡i_{j,n}^{t}italic_i start_POSTSUBSCRIPT italic_j , italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_t end_POSTSUPERSCRIPTacross different convolutional blocks(B)/stages(S)/layers(L) during CTTA. [Top row] - Variations for the domain “glass_blur” across datasets. [Bottom row] - Variations for the domain “snow”. | 2403.10650v4 | 4 |
|
Figure 4:Performance vs. OOD shift.n-MeRCI↓↓\downarrow↓of the UQ methods with increasing OOD shifts in the “easy,” “medium,” and “hard” cases of the GPME.
EnsembleNO andDiverseNOconsistently achieve low n-MeRCI values, indicating higher correlation between their uncertainty estimates and prediction errors. | 2403.10642v2 | 17 |
|
Figure 6:MSE Improvement Ratio forProbConserv.The MSE ratio before and after applyingProbConservto the UQ methods (≥1absent1\geq 1≥ 1indicates improvement) for increasing OOD shifts in the “easy”, “medium” and “hard” cases of the GPME. | 2403.10642v2 | 21 |
|
(b)Coefficient of variation (mean/std) of the first Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 30 |
|
(b)Coefficient of variation (mean/std) of the last Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 33 |
|
(b)Coefficient of variation (mean/std) of the first Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 36 |
|
(b)Coefficient of variation (mean/std) of the last Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 39 |
|
(b)Coefficient of variation (mean/std) of the first Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 42 |
|
(b)Coefficient of variation (mean/std) of the last Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 45 |
|
(b)Coefficient of variation (mean/std) of the first Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 48 |
|
(b)Coefficient of variation (mean/std) of the last Fourier layer weights across the 10 models in the ensemble. | 2403.10642v2 | 51 |
|
Figure 19:Effect ofM𝑀Mitalic_Mon FLOPs.The number of FLOPs for one forward pass as a function of the number of prediction headsM𝑀Mitalic_Mfor various penultimate layer sizes. The diversity regularization strengthλdiversesubscript𝜆diverse\lambda_{\text{diverse}}italic_λ start_POSTSUBSCRIPT diverse end_POSTSUBSCRIPTis fixed at 10.
The number of FLOPs increases mildly with the number of prediction heads. | 2403.10642v2 | 61 |
|
Figure 2:TSDE version of RB-Stein.Nti(s,a)subscriptsuperscript𝑁𝑖𝑡𝑠𝑎N^{i}_{t}(s,a)italic_N start_POSTSUPERSCRIPT italic_i end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ( italic_s , italic_a )is the number of times armi𝑖iitalic_ihas visited(s,a)𝑠𝑎(s,a)( italic_s , italic_a )up to timet𝑡titalic_t.𝒮~tMsubscriptsuperscript~𝒮𝑀𝑡\tilde{\mathcal{S}}^{M}_{t}over~ start_ARG caligraphic_S end_ARG start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTandr~tMsubscriptsuperscript~𝑟𝑀𝑡\tilde{r}^{M}_{t}over~ start_ARG italic_r end_ARG start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTare the state and reward vectors of the topM𝑀Mitalic_Marms observed at timet𝑡titalic_t. | 2403.10638v1 | 2 |
|
Figure 3:Transition Estimation Algorithm Description.df(|S|):=|𝒮|×(|𝒮|−1)assign𝑑𝑓𝑆𝒮𝒮1df(|S|):=|\mathcal{S}|\times(|\mathcal{S}|-1)italic_d italic_f ( | italic_S | ) := | caligraphic_S | × ( | caligraphic_S | - 1 ). | 2403.10638v1 | 3 |
|
Figure 4:RB-Stein with Mirror SVGD and Dynamic Episodes (TS MCR) : Leket. From left to right:N=10𝑁10N=10italic_N = 10,N=50𝑁50N=50italic_N = 50, andN=100𝑁100N=100italic_N = 100arms. | 2403.10638v1 | 4 |
|
Figure 5:RB-Stein with Mirror SVGD and Dynamic Episdoes (TS MCR): Breaking Ground. From left to right:N=10𝑁10N=10italic_N = 10,N=50𝑁50N=50italic_N = 50, andN=100𝑁100N=100italic_N = 100arms. | 2403.10638v1 | 5 |
|
Figure 7:AverageKLD(𝜶^|𝜶⋆)𝐾𝐿𝐷conditional^𝜶superscript𝜶⋆KLD(\hat{\boldsymbol{\alpha}}|\boldsymbol{\alpha}^{\star})italic_K italic_L italic_D ( over^ start_ARG bold_italic_α end_ARG | bold_italic_α start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )using MSVGD using MSVGD withk𝑘kitalic_k-step transition estimation.T∈[50,100,200]𝑇50100200T\in[50,100,200]italic_T ∈ [ 50 , 100 , 200 ],|𝒮|=10𝒮10|\mathcal{S}|=10| caligraphic_S | = 10,k∈[1,5,10]1𝑘15101k\in[1,5,10]1italic_k ∈ [ 1 , 5 , 10 ] 1,ns=40subscript𝑛𝑠40n_{s}=40italic_n start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT = 40. | 2403.10638v1 | 7 |
|
Figure 8:AverageMAE(𝜼^|𝜼⋆)𝑀𝐴𝐸conditional^𝜼superscript𝜼⋆MAE(\hat{\boldsymbol{\eta}}|\boldsymbol{\eta}^{\star})italic_M italic_A italic_E ( over^ start_ARG bold_italic_η end_ARG | bold_italic_η start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT )using MSVGD withk𝑘kitalic_k-step transition estimation.T∈[50,100,200]𝑇50100200T\in[50,100,200]italic_T ∈ [ 50 , 100 , 200 ],|𝒮|=10𝒮10|\mathcal{S}|=10| caligraphic_S | = 10,k∈[1,5,10]1𝑘15101k\in[1,5,10]1italic_k ∈ [ 1 , 5 , 10 ] 1,ns=40subscript𝑛𝑠40n_{s}=40italic_n start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT = 40. | 2403.10638v1 | 8 |
|
Figure 1:(a)Overview of MeDSLIP. MeDSLIP disentangles images and texts into anatomy and pathology streams to construct a fine-grained alignment. Colored boxes in CXR match highlights in the report: red for anatomy and blue for pathology, denoting information from the individual perspective. Variations in line style within the same color indicate detailed cross-perspective information. The combination of colored boxes and text illustrates two causes for entanglement.(b)Image encoding involves disentangling image representations and aligning them with top common text queries. ProtoCL and ICL losses are used for refining alignment, respectively. Existence predictors trained with BCE loss enable the model to make predictions in a zero-shot manner.(c)Triplets are extracted from reports and prompted (Sec.2.1). Then, they are encoded by a text encoder, usually a frozen language model with a learnable linear layer. | 2403.10635v1 | 1 |
|
Figure 2:Comparison between contrastive learning without (a) or with (b) prototypes, using pathology embeddings and anatomy representations as an example. | 2403.10635v1 | 2 |
|
Figure 1:Illustration of the proposed VLM-debias method, designed for improving fairness when sensitive attributes arenotcollected for a dataset. The zero-shot prediction of pre-trained VLMs is used to infer designated sensitive attributes for image samples, with which pseudo attribute groups are clustered (represented in different colors) for each target label. Fairness intervention is applied in a separate target task model training, with the objective to balance group-wise optimization loss by strategic re-sampling. | 2403.10624v2 | 1 |
|
Figure 2:T-SNE visualization of learnt representations of a vanilla blond hair classification model on CelebA dataset. Samples with theblond hairtarget label are shown, and VLMs’ predictions for the sensitive attribute (Att.) w.r.t. true attributes are marked by different colors. Samples that are wrongly predicted by VLMs do not show distinguishable sensitive attribute representations in latent space. Moreover, for the groupMalethat model is unfair towards (inbold), the VLM predicted samples cause major loss/bias. | 2403.10624v2 | 2 |
|
Figure 3:Simulating pseudo sensitive attribute distribution with different correspondences to the true distribution, measured by ARI. Method effectiveness is evaluated by the fairness metric: worst-group accuracy forage𝑎𝑔𝑒ageitalic_a italic_g italic_eandrace𝑟𝑎𝑐𝑒raceitalic_r italic_a italic_c italic_egroup on UTKFace. The method is robust to error and noise in VLM original predictions. | 2403.10624v2 | 3 |
|
Figure 1:Visualizations of the airway OCT segmentation. The first column visualizes OCT frames; the second column compares different teacher modules; the third column compares 2D OCT geometries from the student modules. | 2403.10622v1 | 1 |
|
Figure 1:Long-term Goal: Ultimately, we envision a modular network where different components,pathsπisubscript𝜋𝑖\pi_{i}italic_π start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT, are optimized for different tasks,𝒟jsubscript𝒟𝑗\mathcal{D}_{j}caligraphic_D start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT, each designed by different researchers. The paths, trained on any available hardware type, communicate infrequently across the world, exchanging useful information and enabling new forms of composition. | 2403.10616v1 | 1 |
|
Figure 2:An illustration of the first example from Section2.3. A4444layer neural network, with blockB1subscript𝐵1B_{1}italic_B start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTconsisting of the first2222layers andB2subscript𝐵2B_{2}italic_B start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTconsisting of the next2222layers. Each block has3333choices of module (each with its own parameters), represented by different colors. On the left, we show all of the9999possible paths. On the right, we show a single path. | 2403.10616v1 | 2 |
|
Figure 3:routing More Frequently at Test-Time: At training time (left panel), the router selects the pathπisubscript𝜋𝑖\pi_{i}italic_π start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTusing the prefixz𝑧zitalic_z. We train the chosen path on the whole sequence using the usual language modeling loss. At test time (right panel), the path selected by the router given the prefix is used to score the next chunk of tokens. Then, we re-use the router to choose the most likely path given the new chunk. This process repeats until the whole sequence has been scored. | 2403.10616v1 | 3 |
|
Figure 5:DiPaCo with more capacity: In this example, level 3 modules are path specific, i.e., modules at that level are not shared by paths. | 2403.10616v1 | 5 |
|
Figure 6:Infrastructure. Training workers (inorange) responsible for performing the inner optimization fetch training task from the train task queue (inpurple). Once the training checkpoints are saved to GFS, their paths and metadata (e.g. training step, path ID, phase ID) are written to a Spanner table (inblue) allowing easy look up of checkpoint path given metadata. The evaluation workers and the sharded outer optimization executors (inred) load training checkpoints once they become ready, as signaled by their appearance in the Spanner table. Finally, a monitoring worker (ingreen) periodically checks the health of workers and reboots them if necessary. | 2403.10616v1 | 6 |
|
Figure 7:Sharded Outer Optimization Executor: An example of outer optimization executors sharded by module, which significantly reduces the processing time and memory requirement. | 2403.10616v1 | 7 |
|
Figure 8:Convergence curves of DiPaCo vs Dense Baseline: We first pretrain a150150150150M parameters model for 24k training steps (purple). We then finetune a 16x16 DiPaCoP=256𝑃256P=256italic_P = 256(red). The remaining gap to the dense 1.3B parameters baseline (orange) is reached by gating more frequently at test time. | 2403.10616v1 | 8 |
|
Figure 9:Scaling number of paths in DiPaCo: We report validation perplexity of DiPaCo with different numbers of paths and parameters. We increase capacity by varying both the number of paths (from8to256) and by using path-specific modules. The size of a path (roughly equivalent to serving cost during deployment) is150150150150M parameters for all models in this figure. | 2403.10616v1 | 9 |
|
Figure 10:Generative and discriminative flat MoE models with varying number of paths, three alternating discriminative phases. The “branching” structure of the figure is due to the ancestry of the models: at the far left, all share an initial dense ancestor. At 16K steps, each discriminative-generative pair of experiments shares a generative ancestor, this is continued (restarting the cosine learning rate decay) with the same router (“generative”) or with a sequence of new routers (“discriminative”) | 2403.10616v1 | 10 |
|
Figure 11:Validation perplexity vs number of alternating minimization steps, 16 path flat MoE. As the number of alternating steps increases, the results improve, but each step gives less improvement:14.0→13.38→13.36→13.25→14.013.38→13.36→13.2514.0\rightarrow 13.38\rightarrow 13.36\rightarrow 13.2514.0 → 13.38 → 13.36 → 13.25PPL for phases 0, 1, 2, 3 respectively. All discriminative routing results in the main text use one alternating minimization phase. | 2403.10616v1 | 11 |
|
Figure 2:Model Overview.
To generate lighting-controlled images, we train a light control module similar to[50], conditioned on normal and shading estimation.
We use a custom Residual Control Encoder to encode the control signal for the ControlNet.
Adding a Residual Control Decoder with a reconstruction loss ensures the full control signal is present in the encoded signal. | 2403.10615v2 | 2 |
|
Figure 3:Dataset Generation Pipeline.
We generate a dataset using the Outdoor Laval dataset[12].
We randomly crop images from the panoramas and automatically predict normal, shading, and caption (Sec.3.4).
For our relighting experiments (Sec.3.3), we extend the dataset with relit images using OutCast[8]. | 2403.10615v2 | 3 |
|
Figure 4:Image Synthesis with Consistent Lighting.
Our generated images feature consistent lighting aligned with the target shading for diverse text prompts. | 2403.10615v2 | 4 |
|
Figure 5:In-Domain Image Synthesis with Controllable Lighting.
We can synthesize images under various lighting conditions. | 2403.10615v2 | 5 |
|
Figure 6:Out-of-Domain Image Synthesis with Controllable Lighting.
Our method learns to control the generation process yet maintains the prior of SD[34].
We show various scenes and styles under changing lighting conditions.
The first three images were obtained with estimated normal of real-world images, while for the last two, we used images generated with SDXL[30].
See supplemental for details. | 2403.10615v2 | 6 |
|
Figure 7:Relighting of Real-World Images.
We train a relighting network and evaluate it on real-world images.
Utilizing the diffusion prior helps the generalization to real samples, especially for shading disambiguation. | 2403.10615v2 | 7 |
|
Figure 8:Identity Preservation.
We ablate the effect of our control architecture on relighting.
ControlNet[50]- left - is prone to ignoring part of the control signal, the wall turns reddish and the shadow gets softened.
Our Control Decoder - middle - with control reconstruction loss helps.
Our full residual architecture - right - takes another step and achieves high consistency. | 2403.10615v2 | 8 |
|
Figure 9:Effect of Lighting Representation.
We show that cast shadows provide essential information for the generation process.
We compare our method against a simple N-dot-L shading conditioning, which provides only coarse lighting information to the model, leading to inconsistent lighting with less control. | 2403.10615v2 | 9 |
|
Figure 10:Effect of Normal Conditioning.
Without normal conditioning, it is impossible for the model to infer geometry in the shadowed regions. | 2403.10615v2 | 10 |
|
Figure 11:Effect of Lighting Representation.
We show additional results to our ablation about the effect of cast shadows (Fig.9). | 2403.10615v2 | 11 |
|
Figure 12:Out-of-Domain Image Synthesis Details.
We show the original input image together with the estimated normals and text prompts used to generate our images in (Fig.6). | 2403.10615v2 | 12 |
|
Figure 1:Posterior approximations givenx16subscript𝑥16x_{16}italic_x start_POSTSUBSCRIPT 16 end_POSTSUBSCRIPTby SMC-Wake, Wake, and Defensive Wake. The bottom right panel depicts the exact posterior distribution. | 2403.10610v1 | 1 |
|
Figure 2:Illustration of mass concentration in wake-phase training. Wake-phase reconstructions (middle) of real MNIST digits (top) with label zero all look nearly identical, unlike SMC-Wake reconstructions (bottom). | 2403.10610v1 | 2 |
|
Figure 8:Visualizations of the importance sampling particle approximations withK=1000𝐾1000K=1000italic_K = 1000for wake-phase training. The nine panels are sampled from the inference network at every hundred gradient steps, so at step100,200,…,900100200…900100,200,\dots,900100 , 200 , … , 900. We see that even within the first one thousand gradient steps, mass is concentrated. | 2403.10610v1 | 8 |
|
Figure 1:A comparison between the effects of neuronal deactivation and synaptic pruning as performed on the sentiment analysis model. We observe a gradual decline in performance as more neurons are deactivated. Synaptic Pruning, however, results in the model being relatively unimpaired before being followed by a sharp decline in performance. | 2403.10596v1 | 1 |
|
Figure 3:This graph delineates the change in the sentiment analysis model’s test accuracy over the amount of noise added to the model. We observe a sudden dropoff at a noise std of10−2.5superscript102.510^{-2.5}10 start_POSTSUPERSCRIPT - 2.5 end_POSTSUPERSCRIPTwhich is very similar to the clinical studies results depicted inFig.5. | 2403.10596v1 | 3 |
|
Figure 7:Ablation study comparing the addition noise to different parts of the sentiment analysis model during training. All models are trained over 100 epochs, and model accuracies for the different tests from this ablation study are observable through the sub-figures. | 2403.10596v1 | 5 |
|
Figure 8:Ablation study of adjusting the scale of the Gaussian distribution of the noise, which is only added to specific layers in a model. We find slower rates of decline when noise is added to the dense layer than to the convolutions or to all layers | 2403.10596v1 | 6 |
|
Figure 9:Ablation study comparing the effect of neuronal deactivation with synapse pruning with different methods of noise introduction, one with noise added during training (the upper nine sub-panels), and the other with noise added after training (the lower nine sub-panels). | 2403.10596v1 | 7 |
|
Figure 1:Varying user demands for image processing.Whether a component in an image is considered deteriorative or valuable content hinges on the user’s intention. For a cluttered input image degraded by various components, users might want to (a) remove the interfering fence and reflection, obtaining a rainy street scene, (b) eliminate all degradations to get a clean street scene, (c) keep the watermark for the sake of copyright protection, and (d∼similar-to\sim∼f) extract the reflection, watermark, and fence. | 2403.10520v1 | 1 |
|
Figure 2:Workflow of controllable blind image decomposition (controllable BID).For synthetic input images that exhibit one or multiple degradations, the controllable BID module first predicts what degradations are present. This prediction of components present aids users in making precise instruction prompts. Once these prompts are formulated, the controllable BID module then processes the input image based on these user-provided instructions. The generated images include the restored image, the degradation masks, and images with specific degradations removed. | 2403.10520v1 | 2 |
|
Figure 3:Comparison between our proposed method CBDNet and the state-of-the-art methods on the BID task.The input images have rain streaks, raindrops, snow and haze, and a clean image is the output, which represents the most challenging task in BID. Here, Restormer* denotes the version of Restormer[74]that we trained on the BID dataset with a fixed resolution. Our system not only exhibits a superior image restoration effect but also has a lower number of trainable parameters. | 2403.10520v1 | 3 |
|
Figure 4:Architecture of CBDNet.Upon receiving an input image with various degradation, the encoder transforms it into a deep feature map. This map is then split into several component feature maps by the decomposition block. Each component feature corresponds to a type of degradation. Within the controllablility block, the source classifier utilizes these component feature maps, facilitating source classification and enabling the user to give an instruction prompt easily. This prompt is subsequently converted by the prompt converter into a categorical vector indicating the image processing to be executed. In the recombination block, the chosen component feature maps are mixed with the categorical vector of the prompt and then reconstructed into a restored image by the decoder. | 2403.10520v1 | 4 |
|
Figure 5:Visualization of feature maps output by the decomposition block.The input image is composed of corresponding degradations on the left side, and then produces the feature maps on the right side after passing through the encoder and decomposition block. The red boxes highlight the near-identical nature of feature maps for the same component across different images. In contrast, the blue boxes reveal complete dissimilarities in feature maps for different components. Furthermore, the green boxes illustrate that absent degradations in the input image lead to feature maps without significant patterns, distinct from those with identifiable degradations. | 2403.10520v1 | 5 |
|
Figure 6:Qualitative results of Task I: Multi-weather removal in driving.Row 1-4 represents the Case (3)-(6) in Table1. Our proposed CBDNet demonstrates good performance in the removal of all components. It produces artifact-free results that are aesthetically appealing, whereas all baseline methods tend to leave behind some components, particularly in row 2 and row 4. | 2403.10520v1 | 6 |
|
Figure 8:Qualitative results of Task III.A : Image restoration in multi-degradation removal.Details of all cases are shown in Table3. CBDNet surpasses BIDeN in all cases, effectively restoring images with low visibility. | 2403.10520v1 | 8 |
|
Figure 9:Qualitative results of Task III.B: Degradation masks reconstruction in multi-degradation removel.The input images contains all nine degradations. BIDeN is not able to effectively separate the various degradation components, which is reflected in the failed reconstruction of the degradation masks, such as fence, shadow, and watermark. This also affects the outcome of image restoration, with residues of fence, watermarks, and shadows remaining on the scene image. CBDNet, on the other hand, successfully separates and reconstructs all components. | 2403.10520v1 | 9 |
|
Figure 10:Qualitative results of Task III.C: Controllable blind image decomposition in multi-degradation removal.Degradations of raindrop, haze, snow, and rain streak are presented in the input for Prompts (1)-(3), and all the nine degradations are present in the inputs for Prompts (4)-(6). Prompt (1): remove the haze and the raindrop; Prompt (2): generate the foggy cityscape; Prompt (3): retain the scene image and the snow; Prompt (4): keep the background and all weather components; Prompt (5): compose image with watermark; Prompt (6): restore scene, flare and shadow together; Prompt (7): remove watermark, reflection and shadow. Our method effectively removes the specified degradations as per the prompts, and the output images closely resemble the ideal post-processed versions. | 2403.10520v1 | 10 |
|
Figure 11:Qualitative results of corner cases.Row 1 showcases a scenario in which the user’s prompt includes an error: the instruction to remove haze is invalid since the input image does not contain any haze. Despite this, our method was not misled by the erroneous instruction and accurately executed watermark removal. Row 2 presents a case where the source classifier misidentified the image’s components. Nonetheless, our method adhered to the user’s directive to remove haze correctly. Row 3 reveals the limitations of CBDNet’s restoration ability in heavily obscured areas. While our method effectively removed snow and watermark, it left noticeable distortions in the area previously occupied by the watermark. | 2403.10520v1 | 11 |
|
Figure 12:Qualitative results of Task I.B (Masks reconstruction of multi-weather removal in driving). CBDNet exhibits comparable performance to BIDeN in haze reconstruction and demonstrates superior performance over BIDeN in the reconstruction of rain streak, snow, and raindrop. | 2403.10520v1 | 12 |
Subsets and Splits