image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 14:Performance of AutoGAD in selecting heterogeneous anomaly detectors on selected datasets (results on other datasets are similar and thus omitted). | 2501.14694v1 | 14 |
|
Figure 1:Performance comparison between our modelTAMAand the existing table LLMs on out-of-domain table understanding and general benchmarks. | 2501.14693v1 | 1 |
|
Figure 5:LLaMA 3.1 8B Instruct model’s performance (y-axis) across different numbers of epochs (x-axis).
We fine-tune the model on the 1,500 instruction pairs, with 500 pairs each from FeTaQA, HiTab, and TabFact, at a learning rate of 1.0e-6. | 2501.14693v1 | 22 |
|
Figure 6:Heatmap when we fine-tune the LLaMA 3.1 8B Instruct model on a single dataset (y-axis) and test against the others (x-axis).
In this plot, we fine-tune the model for two epochs at a learning rate of 1.0e-6 with 500 instruction pairs. | 2501.14693v1 | 25 |
|
Figure 7:Heatmap when we fine-tune LLaMA 3.1 8B Instruct model on a single dataset (y-axis) and test against the others (x-axis).
We fine-tune the model for one to three epochs (horizontal directions) at a learning rate of 1.0e-5, 5.0e-6, 1.0e-6, 5.0e-7, 1.0e-7 (vertical direction) with 500 instruction pairs. | 2501.14693v1 | 26 |
|
Figure 8:Heatmap when we fine-tune LLaMA 3.1 8B Instruct model on a single dataset (y-axis) and test against the others (x-axis).
We fine-tune the model for four to six epochs (horizontal directions) at a learning rate of 1.0e-5, 5.0e-6, 1.0e-6, 5.0e-7, 1.0e-7 (vertical direction) with 500 instruction pairs. | 2501.14693v1 | 27 |
|
Figure 1:Architecture of the EYAS fundus image analysis system, illustrating the client-server model, API gateways, analysis microservices for fundus structures, and the report service that synthesizes results for clinical use. | 2501.14689v1 | 1 |
|
Figure 2:The modular analysis pipeline of EYAS, showing the three-step process of localization, segmentation, and classification for fundus characterization. | 2501.14689v1 | 2 |
|
Figure 3:Performance comparison ofONH Shapeclassifiers using different input formats, highlighting the impact of the analysis pipeline steps on classification accuracy. | 2501.14689v1 | 3 |
|
Figure 1:Minimum Angle Subspace Classifier (MASC) accuracy over the layers of the network when the data is projected onto corrupted training subspaces with the indicated corruption degree, for multiple models/datasets. Rows corresponds to plots with the same corruption degree & the columns correspond to the models, as noted. Training accuracy (dashed line) & testing accuracy (dotted line) of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. The number of class-wise PCA components of these models are shown in Figure16in sectionA.9of the Appendix. | 2501.14687v1 | 1 |
|
Figure 2:Minimum Angle Subspace Classifier (MASC) accuracy over the layers of the network when the data set is projected onto corrupted subspace and subspace corresponding to true training labels. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. Training and testing accuracy of the model is shown. FC corresponds to fully connected layer with ReLU activation whereas Flat corresponds to flatten layer without ReLU activation. The respective number of class-wise PCA components for true training label subspaces of the models is shown in Figure17in sectionA.9of the Appendix. | 2501.14687v1 | 2 |
|
Figure 3:Minimum Angle Subspace Classifier (MASC) accuracy over the layers of the generalized network when the data set is projected onto corrupted training subspaces with the indicated corruption degree. Rows corresponds to plots which have the same corruption degree & the columns correspond to the generalized models as noted. Training & testing accuracy of the generalized model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. The respective number of class-wise PCA components of the models is shown in Figure18in SectionA.9of the Appendix. | 2501.14687v1 | 3 |
|
Figure 4:Class-conditioned training data subspaces on layerwise outputs of MLP using PCA. Top: Schematic of MLP model used in the work. Bottom: Creating the class-conditioned training subspace for ReLU (128) layer where 128 are the number of neurons. | 2501.14687v1 | 4 |
|
Figure 5:MASC test accuracy over the layers of the network when the data is projected onto corrupted training subspaces with the indicated corruption degree. Best Model Testing Accuracy corresponds accuracy of the testing data of the model if early stopping was used. | 2501.14687v1 | 5 |
|
Figure 6:MASC test accuracy over the layers of the network when the data set is projected onto subspace corresponding to true training labels. Best Model Testing Accuracy corresponds accuracy of the testing data of the model if early stopping was used. | 2501.14687v1 | 6 |
|
Figure 7:MASC test accuracy over the layers of AlexNet trained on Tiny ImageNet when the data is projected onto corrupted training subspaces with the indicated corruption degree. Testing accuracy of the model and best model testing accuracy is shown for comparison. Best Model Testing Accuracy corresponds accuracy of the testing data of the model if early stopping was used. | 2501.14687v1 | 7 |
|
Figure 8:MASC test accuracy over the layers of AlexNet trained on Tiny ImageNet when the data set is projected onto corrupted training and true training subspace. Testing accuracy of the model and best model testing accuracy is shown for comparison. Best Model Testing Accuracy corresponds accuracy of the testing data of the model if early stopping was used. | 2501.14687v1 | 8 |
|
Figure 9:MASC accuracy over the layers of trained and random initialized network when the data is projected onto corrupted training subspaces with the indicated corruption degree. | 2501.14687v1 | 9 |
|
Figure 10:MASC accuracy over the layers of trained and random initialized network when the data set is projected onto subspace corresponding to true training labels. Testing accuracy of the trained model is shown for comparison. | 2501.14687v1 | 10 |
|
Figure 11:MASC accuracy over the layers of trained and random initialized AlexNet-Tiny ImageNet when the data is projected onto corrupted training subspaces with the indicated corruption degree. | 2501.14687v1 | 11 |
|
Figure 12:MASC accuracy over the layers of trained and random initialized AlexNet-Tiny ImageNet when the data set is projected onto subspace corresponding to true training labels. | 2501.14687v1 | 12 |
|
Figure 13:MASC accuracy over the layers of the network when the data is projected onto corrupted training subspaces with the indicated corruption degree. The number of class-wise PCA components of these models are shown in Figure16in sectionA.9of the Appendix. | 2501.14687v1 | 13 |
|
Figure 14:MASC accuracy over the layers of the network when the data set is projected subspace corresponding to true training labels. The respective number of class-wise PCA components for true training label subspaces of the models is shown in Figure17in sectionA.9of the Appendix. | 2501.14687v1 | 14 |
|
Figure 15:MASC accuracy over the layers of the generalized network when the data set is projected onto corrupted training subspaces with the indicated corruption degree. The respective number of class-wise PCA components of the models is shown in Figure18in SectionA.9of the Appendix. | 2501.14687v1 | 15 |
|
Figure 16:Class-wise number of PCA components of the corrupted training subspace used in section4over the layers of multiple networks with various corruptions degrees. Although it is not mentioned in the legend, all the 100 classes and 200 classes of CIFAR-100 and Tiny ImageNet respectively are plotted. | 2501.14687v1 | 16 |
|
Figure 17:Class-wise number of PCA components of the subspace corresponding to true training labels used in section5over the layers of multiple networks with various corruptions. Although it is not mentioned in the legend, all the 100 classes and 200 classes of CIFAR-100 and Tiny ImageNet respectively are plotted. | 2501.14687v1 | 17 |
|
Figure 18:Class-wise number of PCA components of the corrupted training subspace used in section6over the layers of multiple generalized networks with various corruption degrees. Although it is not mentioned in the legend, all the 100 classes and 200 classes of CIFAR-100 and Tiny ImageNet respectively are plotted. | 2501.14687v1 | 18 |
|
Figure 19:MASC accuracy over the layers of the MLP network when the data is projected onto corrupted
training subspaces with the indicated corruption degree, for MLP models with MNIST and CIFAR10 datasets. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_moptimizer as noted. Training and testing accuracy of the model is shown.FC corresponds to fully connected layer with ReLU activation. | 2501.14687v1 | 19 |
|
Figure 20:Class-wise number of PCA components of the corrupted training subspace over the layers of MLP networks trained with MNIST and CIFAR10 datasets with various corruption degree. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_moptimizer as noted. | 2501.14687v1 | 20 |
|
Figure 21:MASC accuracy over the layers of the MLP network when the data set is projected onto subspace
corresponding to true training labels. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_moptimizer as noted. Training and testing accuracy of the model
is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 21 |
|
Figure 22:Class-wise number of PCA components of the subspace corresponding to true training
labels over the layers of MLP networks with various corruption degrees. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_moptimizer as noted. | 2501.14687v1 | 22 |
|
Figure 23:MASC accuracy over the layers of the generalized MLP network when the data set is projected onto corrupted training subspaces with the indicated corruption degree. Rows corresponds to plots which have the same corruption degree & the columns correspond to the generalized models withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_mas noted. Training & testing accuracy of the generalized model withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_mis shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 23 |
|
Figure 24:Class-wise number of PCA components of the corrupted training subspace over the layers of generalized MLP network with various corruption degrees. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models withSGD𝑆𝐺𝐷SGDitalic_S italic_G italic_DandAdam𝐴𝑑𝑎𝑚Adamitalic_A italic_d italic_a italic_moptimizer as noted. | 2501.14687v1 | 24 |
|
Figure 25:MASC accuracy over the layers of the network when the data is projected onto corrupted
training subspaces with the indicated corruption degree, for multiple CNN models trained with and without dropout. Rows corresponds to plots with the same corruption degree & the columns correspond to the models, as noted.
Training accuracy (dashed line) & testing accuracy (dotted line) of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 25 |
|
Figure 26:Class-wise number of PCA components of the corrupted training subspace over the layers of CNN networks trained with and without dropout, for various corruption degrees. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. | 2501.14687v1 | 26 |
|
Figure 27:MASC accuracy over the layers of the network when the data is projected onto corrupted
training subspaces with the indicated corruption degree, for CNN and AlexNet models trained with and without dropout. Rows corresponds to plots with the same corruption degree & the columns correspond to the models as noted.
Training accuracy (dashed line) & testing accuracy (dotted line) of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation hereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 27 |
|
Figure 28:Class-wise number of PCA components of the corrupted training subspace over the layers of CNN and AlexNet networks trained with and without dropout, for various corruption degrees. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. | 2501.14687v1 | 28 |
|
Figure 29:MASC accuracy over the layers of the CNN network, trained with and without dropout, when the data set is projected onto subspace corresponding to true training labels. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. Training and testing accuracy of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 29 |
|
Figure 30:Class-wise number of PCA components of the subspace corresponding to true training
labels over the layers of CNN networks trained with and without drop out and various corruption degrees. Although it is
not mentioned in the legend, all the 100 classes of CIFAR-100 are plotted. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. | 2501.14687v1 | 30 |
|
Figure 31:MASC accuracy over the layers of the CNN and AlexNet, trained with and without dropout, when the data set is projected onto subspace corresponding to true training labels. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. Training and testing accuracy of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 31 |
|
Figure 32:Class-wise number of PCA components of the subspace corresponding to true training
labels over the layers of CNN and AlexNet models trained with and without drop out and various corruption degrees. Although it is not mentioned in the legend, all the 100 classes of CIFAR-100 are plotted. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. | 2501.14687v1 | 32 |
|
Figure 33:MASC accuracy over the layers of the generalized CNN network, trained with and without drop out, when the data set is projected onto corrupted training subspaces with the indicated corruption degree. Rows corresponds to plots
which have the same corruption degree & the columns correspond to the generalized models as
noted. Training & testing accuracy of the generalized model is shown. FC corresponds to fully
connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 33 |
|
Figure 34:Class-wise number of PCA components of the corrupted training subspace over the layers of generalized CNN networks with various corruption degrees. Although it is not mentioned in the legend, all the 100 classes of CIFAR-100 are plotted. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models with and without drop out as noted. | 2501.14687v1 | 34 |
|
Figure 35:MASC accuracy over the layers of the generalized CNN and AlexNet network, when the data set is projected onto corrupted training subspaces with the indicated corruption degree. Rows corresponds to plots which have the same corruption degree & the columns correspond to the generalized models trained with and without drop out as noted. Training & testing accuracy of the generalized model is shown. FC corresponds to fully
connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 35 |
|
Figure 36:Class-wise number of PCA components of the corrupted training subspace over the layers of generalized CNN and AlexNet networks with various corruption degrees. Although it is not mentioned in the legend, all the 100 classes of CIFAR-100 are plotted. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models with and without drop out as noted. | 2501.14687v1 | 36 |
|
Figure 37:MASC accuracy over the layers of the AlexNet model trained on Tiny ImageNet when the data is projected onto 99% and 90% variance explained corrupted training subspaces with the indicated corruption degree, for multiple models/datasets. Rows corresponds to plots with the same corruption degree & the columns correspond to the models, as noted. Training accuracy (dashed line) & testing accuracy (dotted line) of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 37 |
|
Figure 38:MASC accuracy over the layers of the AlexNet models trained on Tiny ImageNet when the data set is projected onto 99.9%, 99% and 90% variance explained subspace corresponding to true training labels. Rows corresponds to plots which have the same corruption degree and the columns correspond to the models as noted. Training and testing accuracy of the model is shown. FC corresponds to fully connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds
to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 38 |
|
Figure 39:MASC accuracy over the layers of the generalized AlexNet model trained on Tiny ImageNet when the data set is projected onto 99% and 90% variance explained training subspaces with the indicated corruption degrees. Rows corresponds to plots which have the same corruption degree & the columns correspond to the generalized models as
noted. Training & testing accuracy of the generalized model is shown. FC corresponds to fully
connected layer withReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation whereas Flat corresponds to flatten layer withoutReLU𝑅𝑒𝐿𝑈ReLUitalic_R italic_e italic_L italic_Uactivation. | 2501.14687v1 | 39 |
|
Figure 1:Framework of our study. We evaluate the performance using the MedMNIST dataset collection and select foundation models from a representative pool. | 2501.14685v1 | 1 |
|
Figure 2:Accuracy of DINO ViT-B/16 on DermaMNIST with various numbers of training data for each class, and “full” means using all training data. | 2501.14685v1 | 2 |
|
Figure 3:Comparing the Accuracy of VGG16 with the learning rate of the encoder ranging in{10−3,10−4,10−5}superscript103superscript104superscript105\{10^{-3},10^{-4},10^{-5}\}{ 10 start_POSTSUPERSCRIPT - 3 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 3 |
|
Figure 4:Comparing the Accuracy of ResNet-18 with the learning rate of the encoder ranging in{10−3,10−4,10−5}superscript103superscript104superscript105\{10^{-3},10^{-4},10^{-5}\}{ 10 start_POSTSUPERSCRIPT - 3 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 4 |
|
Figure 5:Comparing the Accuracy of DenseNet-121 with the learning rate of the encoder ranging in{10−3,10−4,10−5}superscript103superscript104superscript105\{10^{-3},10^{-4},10^{-5}\}{ 10 start_POSTSUPERSCRIPT - 3 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 5 |
|
Figure 6:Comparing the Accuracy of EfficientNet-B4 with the learning rate of the encoder ranging in{10−3,10−4,10−5}superscript103superscript104superscript105\{10^{-3},10^{-4},10^{-5}\}{ 10 start_POSTSUPERSCRIPT - 3 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 6 |
|
Figure 7:Tuning the learning rate for the encoder of ViT-B/16 ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 7 |
|
Figure 8:Comparing the Accuracy of CLIP ViT-B/16 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 8 |
|
Figure 9:Comparing the Accuracy of RSAM-C ViT-B/16 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 9 |
|
Figure 10:Comparing the Accuracy of SAM ViT-B/16 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 10 |
|
Figure 11:Comparing the Accuracy of EVA-02 ViT-B/14 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 11 |
|
Figure 12:Comparing the Accuracy of OpenAI ViT-B/16 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 12 |
|
Figure 13:Comparing the Accuracy of DINO ViT-B/16 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 13 |
|
Figure 14:Comparing the Accuracy of DINOv2 ViT-B/14 with the learning rate of the encoder ranging in{10−4,10−5,10−6}superscript104superscript105superscript106\{10^{-4},10^{-5},10^{-6}\}{ 10 start_POSTSUPERSCRIPT - 4 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 5 end_POSTSUPERSCRIPT , 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCRIPT }on different datasets. | 2501.14685v1 | 14 |
|
Figure 1:Remote Robotic Surgery Framework Utilizing TI and Informer Model for Enhanced PSM Precision | 2501.14678v1 | 1 |
|
Figure 3:The top plot (Part 1) shows the simulated packet loss pattern across 1000 time steps. The following plots (Parts 2-4) display the original and corrupted tool tip position along all three axes, with gray-shaded regions highlighting periods of packet loss. | 2501.14678v1 | 3 |
|
Figure 4:Prediction performance of the Informer model under packet loss for tool tip position in X, Y, and Z axes. Solid and dashed lines represent actual and predicted positions, respectively. The model achieves accuracies of 96.68%, 95.96%, and 90.37% for the X, Y, and Z axes, demonstrating robustness against network-induced packet loss. | 2501.14678v1 | 4 |
|
Figure 2:Definitions and motivations for MatAnyone.
(a) In a matting frame, the image can be broadly divided into two areas based on the alpha value: thecore(semantic) and theboundary(fine-details). Thecoreincludes the background (alpha values of 0) and the solid foreground (alpha values of 1), while theboundary(highlighted in pink) encompasses areas with alpha values between 0 and 1.
(b) Due to the under-defined setting, auxiliary-free methods like RVM[33]are easily confused by ambiguous background. Meanwhile, mask-guided methods like MaGGIe[22]tend to break the segmentation prior they aim to leverage, due to the deficiency in video matting data. | 2501.14677v1 | 2 |
|
Figure 3:An overview of MatAnyone.
MatAnyone is a memory-based framework for video matting. Given a target segmentation map in the first frame, our model achieves stable and high-quality matting through consistent memory propagation, with a region-adaptive memory fusion module to combine information from the previous and current frame. To overcome the scarcity of real video matting data, we incorporate a new training strategy that effectively leverages matting data for fine-grained matting details and segmentation data for semantic stability, with designed losses separately. | 2501.14677v1 | 3 |
|
Figure 4:Qualitative comparisons on real-world videos.
Our MatAnyone significantly outperforms existing auxiliary-free (RVM[33]) and mask-guided (FTP-VM[21]and MaGGIe[22]) approaches in both detail extraction and semantic accuracy. For the lowest row, while other methods all miss out on important body parts (i.e., head) and mistakenly take background pixels as foreground (due to similar colors), thus generating messy outputs, our method presents an accurate and visually clean output by even identifying the shadow near the boundary. | 2501.14677v1 | 4 |
|
Figure 5:Quantitative comparisons with MaGGIe[22]on instance video matting. Despite MaGGIe using instance mask as guidance for each frame, our method shows better performance, achieving better stability in object tracking and finer alpha matte details. | 2501.14677v1 | 5 |
|
Figure 7:Comparison of matting results training with original DDC loss[35]and with scaled DDC loss, where the latter gives more stable and natural matting results. | 2501.14677v1 | 7 |
|
Figure 8:Issues with VideoMatte240K[32].(a) Errors in alpha values exist in reflective regions (e.g., “a hole” on glasses).
(b) Inhomogeneous alpha values exist in core regions (e.g., caused by shadow), where the alpha value should be exactly 0 or 1. | 2501.14677v1 | 8 |
|
Figure 9:Gallery for our new training dataset VM800.High-quality details in the boundary regions and diversity in terms of gender, hairstyles, and aspect ratios could be clearly observed. | 2501.14677v1 | 9 |
|
Figure 10:Harmonization on synthetic benchmarks and its effect on model performance.Harmonization[23]is an operation that makes the composited frame more natural and realistic, which also effectively makes our YouTubeMatte a more challenging benchmark that is closer to the real distribution.
It is observed that while RVM[33]is confused by the harmonized frame, our method still yields robust performance. | 2501.14677v1 | 10 |
|
Figure 11:(a) Comparison on results trained with old training data (VideoMatte240K[32]) and new training data (our VM800).It could be observed that training with old data will lead to errors in reflective objects (e.g., holes on the sunglasses) and inhomogeneous alpha values in the core regions. However, both issues are fixed when training with our new data, indicating a higher quality.(b) Comparison on results trained without and with core-area supervision.It could be observed that training without it will lead to semantics error due to the weak supervision from real segmentation data, while training with core supervision largely improves semantics accuracy thanks to the stronger supervision enabled. | 2501.14677v1 | 11 |
|
Figure 12:Comparison on results with and without Consistent Memory Propagation.It could be observed that when CMP is not applied, semantic errors constantly exist across a wide span of video frames. However, when training with CMP, we observe from the “Change Probability” mask that usually our model only takes pixels near the boundary as “changed”, and most of the inner regions (i.e., earring) will mainly take the memory values from the last frame. As we can see on the figure, while predictions are both correct at timet𝑡titalic_t, the model with CMP successfully keeps the correctness and gives stable results, while the model without CMP quickly breaks the correctness and never recovers. | 2501.14677v1 | 12 |
|
Figure 13:Comparison on results with iterative refinement.A noticeable enhancement on details can be observed even with one iteration of refinement compared with the given segmentation mask. Within 10 iterations, our model is able to achieve matting details at an image-matting level, even better than Matte Anything[56], which is an image matting model also based on the results from SAM[25]. | 2501.14677v1 | 13 |
|
Figure 14:More qualitative comparisons on general video matting with SOTA methods.We compare our MatAnyone with both auxiliary-free (AF) method: RVM[33]and mask-guided methods: FTP-VM[21], and MaGGIe[22].
It could be observed that our method significantly outperforms others in both detail extraction and semantic accuracy, across diverse and complex real scenarios.
It is noteworthy that although sometimes MaGGIe[22]seems to give acceptable results when compositing with a green screen, its alpha matte turns out to be noisy (i.e., inhomogeneous in the core foreground region and blurry in the boundary region), while our alpha matte is clean with fine-grained details in the boundary region. As a result, we also include alpha mattes for a more comprehensive comparison.(Zoom in for best view) | 2501.14677v1 | 14 |
|
Figure 15:A challenging example of general video matting across a long time span.We compare our MatAnyone with both auxiliary-free (AF) method: RVM[33]and mask-guided methods: FTP-VM[21], and MaGGIe[22].
It could be observed that our model is able to track the target object stably even when the object is moving fast in a highly complex scene, where all the other methods present noticeable failures.(Zoom in for best view) | 2501.14677v1 | 15 |
|
Figure 16:Another challenging example of general video matting across a long time span.We compare our MatAnyone with both auxiliary-free (AF) method: RVM[33]and mask-guided methods: FTP-VM[21], and MaGGIe[22].
This example showcases that our model is able to track the target objects even in a highly ambiguous background, where the colors for foreground and background are similar, and also multiple humans in the background.
In addition, it also demonstrates when there is more than one target object, our model is still able to handle this challenging case well.(Zoom in for best view) | 2501.14677v1 | 16 |
|
Figure 17:More qualitative comparisons on instance matting.We compare our MatAnyone with MaGGIe[22], a mask-guided method that requires the instance mask foreachframe, while our method only requires the mask for thefirstframe.
It could be observed that even with such strong given prior, MaGGIe still performs below our method in terms of semantic accuracy in the core regions.
Moreover, in terms of the boundary regions, by examining the details there, we could clearly observe that the details generated by MaGGIe are blurry and far from fine-grained compared with our results.(Zoom in for best view) | 2501.14677v1 | 17 |
|
Figure 2:Experimental setup. The Xilinx RFSoC ZCU216 runs on custom QICK firmware and software for qubit control and readout. The qubit readout and drive pulses are combined before entering the fridge; the flux pulse is used for qubit cooling. The readout signal is amplified before being digitalized by the ADC. | 2501.14663v1 | 2 |
|
Figure 3:Typical single shot readout signal. The readout phase is adjusted to optimize the signal inQquadrature. (a) Single-shot readout trajectory and averaged readout trajectory. (b)(c) Time-averaged single-shot readout signal in the I-Q plane and its histogram (showing 5000 shots for better visibility). The gray dashed line represents the readout threshold for optimized readout fidelity. | 2501.14663v1 | 3 |
|
Figure 1:Challenging lighting environment and the texture of the object’s surface can cause errors in the SL encoding information. The algorithm we propose can effectively reduce such errors. | 2501.14659v1 | 1 |
|
Figure 4:In Scene 3, the EC method incorrectly changes the exposure, leading to a significant increase in the decoding error rate. | 2501.14659v1 | 4 |
|
Figure 5:The comparison of the optimized decoding error rates of different methods under various scenes is shown, where the left figure presents the optimization results for binary stripes (Gray code), and the right figure shows the optimization results for speckle structured light. | 2501.14659v1 | 5 |
|
Figure 8:Global matching results with and without the error code (EC) filter. The fourth column is an enlarged view of the erroneous matches, with red dots indicating incorrect matches, green dots indicating correct matches, and blue dots indicating unmatched points. | 2501.14659v1 | 10 |
|
Figure 1:Illustrative toy task on two settings. 1) FL settings, where all users are participating in the training (left figure), 2) FDG setting: one user is excluded in the training (right figure). | 2501.14653v1 | 1 |
|
Figure 2:Performance comparison without pretrained models forα=1,U=20formulae-sequence𝛼1𝑈20\alpha=1,U=20italic_α = 1 , italic_U = 20. | 2501.14653v1 | 2 |
|
Figure 9:Performance comparison of FL algorithms without pretrained models forα=0.1,U=20formulae-sequence𝛼0.1𝑈20\alpha=0.1,U=20italic_α = 0.1 , italic_U = 20. | 2501.14653v1 | 13 |
|
Figure 10:Performance comparison of FL algorithms without pretrained models forα=0.1,U=40formulae-sequence𝛼0.1𝑈40\alpha=0.1,U=40italic_α = 0.1 , italic_U = 40. | 2501.14653v1 | 14 |
|
Figure 11:Performance comparison of FL algorithms without pretrained models forα=1,U=40formulae-sequence𝛼1𝑈40\alpha=1,U=40italic_α = 1 , italic_U = 40. | 2501.14653v1 | 15 |
|
Figure 12:Performance comparison of FL algorithms without pretrained models forα=0.1,U=60formulae-sequence𝛼0.1𝑈60\alpha=0.1,U=60italic_α = 0.1 , italic_U = 60. | 2501.14653v1 | 16 |
|
Figure 13:Performance comparison of FL algorithms without pretrained models forα=1,U=60formulae-sequence𝛼1𝑈60\alpha=1,U=60italic_α = 1 , italic_U = 60. | 2501.14653v1 | 17 |
|
Figure 1:Trajectories (top row) and distance to equilibrium over synchronization rounds (bottom row) of GDA (K=1𝐾1K=1italic_K = 1) and Decoupled SGDA withK={2,5}𝐾25K=\{2,5\}italic_K = { 2 , 5 }on the (QG) problem (d=2𝑑2d=2italic_d = 2).𝐂𝐂\mathbf{C}bold_Cin (QG) is a constant here—the larger, the stronger the interactive term.Left-to-right:decreasing the constantc∈{10,3.5,2,7,0}𝑐103.5270c\in\{10,3.5,2,7,0\}italic_c ∈ { 10 , 3.5 , 2 , 7 , 0 }.
The markers denote the local steps and star the solution.
See §5for discussion. | 2501.14652v1 | 1 |
|
Figure 2:Number of rounds (log-scale; lower is better) to reach epsilon accuracy for varyingλmax(𝐂)subscript𝜆max𝐂\lambda_{\text{max}}(\mathbf{C})italic_λ start_POSTSUBSCRIPT max end_POSTSUBSCRIPT ( bold_C )in (QG).Left:Decoupled GDA with differentK𝐾Kitalic_K-values and GDA (K=1𝐾1K=1italic_K = 1).Right:comparison between GDA, Decoupled GDA, Optimistic GDA(Popov,1980), ALT–alternating GDA and Extragradient(Korpelevich,1976). | 2501.14652v1 | 2 |
|
Figure 3:Lowest gradient norm reached after a fixed number of communication rounds, for varying1/λ1𝜆1/\lambda1 / italic_λin (toyGAN).Left:Effect ofK𝐾Kitalic_K.Right:different optimization methods, GDA, Decoupled GDA, Optimistic GDA(Popov,1980), ALT–alternating GDA and Extragradient(Korpelevich,1976).
See §5for discussion. | 2501.14652v1 | 3 |
|
Figure 4:Lowest gradient norm reached by Decoupled SGDA and Local SGDA for a fixed number of communication rounds using unbalanced noisy gradient oracles.Left:Decoupled SGDAvs.Federated Minimaxfor varying values of∥𝐂∥delimited-∥∥𝐂\left\lVert\mathbf{C}\right\rVert∥ bold_C ∥and fixed variance.Left:Decoupled SGDAvs.Local SGDAfor varying levels of off-diagonal variance noise (σuvsubscript𝜎𝑢𝑣\sigma_{uv}italic_σ start_POSTSUBSCRIPT italic_u italic_v end_POSTSUBSCRIPT,σvusubscript𝜎𝑣𝑢\sigma_{vu}italic_σ start_POSTSUBSCRIPT italic_v italic_u end_POSTSUBSCRIPT). See5. | 2501.14652v1 | 4 |
|
Figure 5:y𝑦yitalic_y-axis: FID scores (log scale; lower is better) during GAN training, versusx𝑥xitalic_x-axis communication rounds.Left:results on theCIFAR-10(Krizhevsky,2009)dataset.Right:results on theSVHN(Netzer et al.,2011)dataset. | 2501.14652v1 | 5 |
|
Figure 7:Comparison of our method with the federated minimax formulation: Our method splits the parameter space, while the federated formulation splits the data. Moreover, our method only allows each player to access the gradient with respect to their own parameters, whereas in federated minimax, each player can compute the gradient with respect to both their own parameters and the other player’s parameters. | 2501.14652v1 | 7 |
|
Figure 8:Trajectories and convergence comparison of GDA,Decoupled SGDAandDecoupled SGDA with Ghost Sequencewith different values of𝐂=C𝐈𝐂𝐶𝐈\mathbf{C}=C\mathbf{I}bold_C = italic_C bold_I(interaction strength). The top row shows the trajectories of the different algorithms forK={1,5}𝐾15K=\{1,5\}italic_K = { 1 , 5 }over varying values ofC∈{25,15,5,0.1}𝐶251550.1C\in\{25,15,5,0.1\}italic_C ∈ { 25 , 15 , 5 , 0.1 }. AsC𝐶Citalic_Cdecreases, trajectories become more stable, with the Decoupled SGDA with Ghost Sequence (blue) showing more efficient convergence compared to GDA (black) and Decoupled SGDA (red). The bottom row presents the synchronization rounds versus distance to equilibrium for each configuration, highlighting faster convergence of Decoupled SGDA with Ghost Sequence under larger𝐂𝐂\mathbf{C}bold_Cvalues, while Decoupled SGDA with Ghost Sequence and Decoupled SGDA converge similarly for small𝐂𝐂\mathbf{C}bold_C. | 2501.14652v1 | 8 |
Subsets and Splits