🎯 What it does: The concept of first-order flatness is proposed, and based on this, the Gradient Norm Aware Minimization (GAM) algorithm is designed to explicitly optimize the gradient norm of model weights during training, thereby seeking flatter minima and enhancing the model's generalization performance.
🎯 What it does: This paper proposes GradMA, an accelerated federated learning framework that simultaneously utilizes gradient memory on both the server and client sides, aimed at alleviating catastrophic forgetting caused by data heterogeneity and partial participation.
🎯 What it does: A completely unsupervised 3D point cloud semantic segmentation method called GrowSP is proposed, which achieves point-level semantic segmentation using progressively expanded superpoints and semantic primitive clustering.
🎯 What it does: A guided deep super-resolution method that integrates deep learning with anisotropic diffusion is proposed, utilizing RGB guidance images to achieve high-quality depth image magnification.
Guiding Pseudo-Labels With Uncertainty Estimation for Source-Free Unsupervised Domain Adaptation
Mattia Litrico (Istituto Italiano di Tecnologia), Pietro Morerio (Istituto Italiano di Tecnologia)
CodeDomain AdaptationContrastive LearningImage
🎯 What it does: A source-free unsupervised domain adaptation method is proposed, which enhances target domain performance by re-weighting pseudo-labels through neighbor knowledge aggregation and entropy estimation.
🎯 What it does: This paper proposes enhancing contrastive learning in self-supervised learning for skeleton action recognition by synthesizing hard positive samples (HaLP) in the latent space.
🎯 What it does: A few-shot handwritten text generation model based on Transformer, VATr, is proposed, which utilizes visual prototypes to encode and generate images of a specific author's writing style.
🎯 What it does: This work presents Handy, a high-fidelity hand shape and texture parameter model based on over 1200 hand scans, capable of reconstructing 3D hand shapes and high-frequency textures from a single photo.
🎯 What it does: This paper proposes a Hard Patches Mining (HPM) framework, where the model acts as both student and teacher during MIM pre-training, generating more challenging occlusion tasks by predicting the reconstruction loss of each patch, and jointly training the reconstruction network and loss predictor.
🎯 What it does: This paper addresses the issue of performance degradation caused by the easy fitting of synthetic samples in zero-shot quantization, proposing the Hard Sample Synthesis and Training (HAST) method.
🎯 What it does: This paper proposes a new Harmonious Feature Learning Network (HFL-Net) for simultaneously estimating the 3D poses of hands and objects from a single RGB image.
Harmonious Teacher for Cross-Domain Object Detection
Jinhong Deng (University of Electronic Science and Technology of China), Lixin Duan (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China)
🎯 What it does: Proposes the Harmonious Teacher framework, which enhances the classification and localization consistency of detection models in the source and target domains through self-supervised and unsupervised harmonious loss, and improves the self-training method for cross-domain object detection by using harmonious metrics for threshold-free weighting of pseudo-labels.
🎯 What it does: For surface reconstruction of multi-view indoor scenes, the HelixSurf method is proposed, which combines traditional PatchMatch MVS with neural implicit surface learning, and enhances geometric details and robustness through mutually iterative regularization.
Hidden Gems: 4D Radar Scene Flow Learning Using Cross-Modal Supervision
Fangqiang Ding (University of Edinburgh), Chris Xiaoxuan Lu (Delft University of Technology)
CodeAutonomous DrivingSimultaneous Localization and MappingOptical FlowMultimodalityPoint Cloud
🎯 What it does: This paper proposes a 4D radar scene flow estimation method based on multi-modal collaborative supervision, using co-located radar, LiDAR, camera, and odometry data for training without manual labeling.
🎯 What it does: Using a teacher-student framework, the pseudo-labels output by the teacher network are dynamically divided into three groups: high confidence, ambiguous, and low confidence. Shuffle data augmentation is introduced in the student network to enhance feature representation capabilities.
Hierarchical Video-Moment Retrieval and Step-Captioning
Abhay Zala (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
CodeGenerationRetrievalTransformerSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: The HIREST dataset and a four-level video retrieval and step generation task are proposed, constructing an end-to-end retrieval and decomposition framework.
🎯 What it does: A scalable frequency domain decomposition network is designed and implemented to achieve high-fidelity 3D hand model reconstruction from a single image.
🎯 What it does: A high-fidelity 3D GAN inversion framework based on pseudo-multi-view optimization is proposed, capable of generating high-quality, 3D-consistent view synthesis from a single image, and supports attribute editing and texture modification.
🎯 What it does: This paper proposes a full-body 3D human digitization framework (2K2K) based on a single 2K (2048×2048) high-resolution color image. It generates a high-fidelity 3D mesh through segmentation (body parts) extraction, surface normal prediction, low-resolution full-body depth prediction, and high-resolution depth fusion.
High-Fidelity Clothed Avatar Reconstruction From a Single Image
Tingting Liao (University of Chinese Academy of Sciences), Zhen Lei (University of Chinese Academy of Sciences)
CodeGenerationPose EstimationImage
🎯 What it does: This paper proposes a coarse-to-fine two-stage framework for reconstructing human avatars from a single image (CAR), enabling the rapid generation of high-fidelity clothing pose avatars.
Histopathology Whole Slide Image Analysis With Heterogeneous Graph Representation Learning
Tsai Hor Chan (University of Hong Kong), Lequan Yu (University of Hong Kong)
CodeClassificationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper constructs a heterogeneous graph that includes cell types and continuous similarity, utilizing the HEAT layer and pseudo-label pooling for whole slide image (WSI) analysis, and provides causal explanations.
HOICLIP: Efficient Knowledge Transfer for HOI Detection With Vision-Language Models
Shan Ning (ShanghaiTech University), Xuming He (ShanghaiTech University)
CodeObject DetectionKnowledge DistillationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes the HOICLIP framework, which efficiently transfers CLIP's visual-language knowledge to the human-object interaction (HOI) detection task.
🎯 What it does: Improving video action recognition using external object detection information, proposing Object-Guided Token Sampling (OGS) and Object-Aware Attention Module (OAM) to achieve the dual goals of reducing token inefficiency and enhancing accuracy.
🎯 What it does: This paper proposes a backdoor attack framework for diffusion models called BadDiffusion, and demonstrates the feasibility of this attack in image generation tasks.
🎯 What it does: This paper proposes a new hybrid scale feature extraction layer (HS-layer) and applies it to the category-level object pose estimation framework HS-Pose, which can simultaneously capture local and global geometric information, encode scale and translation information, and is robust to outliers.
Hubs and Hyperspheres: Reducing Hubness and Improving Transductive Few-Shot Learning With Hyperspherical Embeddings
Daniel J. Trosten (UiT Arctic University of Norway), Michael C. Kampffmeyer (UiT Arctic University of Norway)
CodeClassificationRepresentation LearningImage
🎯 What it does: This paper proposes a method to eliminate the hubness problem in transductive few-shot learning by achieving uniform embedding on the sphere, enhancing classification performance while maintaining local similarity (LSP);
🎯 What it does: A multi-version high-quality GT generation process based on human judgment has been developed, and a specialized HGGT dataset for super-resolution in real scenarios has been constructed based on this.
Zigang Geng (University of Science and Technology of China), Han Hu (University of Science and Technology of China)
CodePose EstimationTransformerImage
🎯 What it does: Proposes a structured representation called Pose as Compositional Tokens (PCT), which maps human poses to discrete sub-structure tokens and completes pose estimation through classification tasks.
🎯 What it does: This paper presents the Human-Art dataset, aimed at bridging the gap between natural scenes and artificial scenes (such as sculptures, paintings, cartoons, digital art, etc.) in the tasks of human detection and pose estimation, and provides rich annotations (bounding boxes, 21 key points, self-contact points, text descriptions) to support various downstream tasks.
🎯 What it does: This paper proposes HumanBench (which includes 19 datasets covering 6 categories of human perception tasks) and PATH (a projector-based hierarchical weight sharing pre-training method) for learning general human visual representations.
Hunting Sparsity: Density-Guided Contrastive Learning for Semi-Supervised Semantic Segmentation
Xiaoyang Wang (University of Liverpool), Jimin Xiao (XJTLU)
CodeSegmentationContrastive LearningImage
🎯 What it does: This paper proposes a density-guided contrastive learning framework based on feature space (DGCL), which enhances semi-supervised semantic segmentation by locating sparse features and guiding them to cluster around high-density centers.
CodePose EstimationAutonomous DrivingSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: This paper proposes a new LiDAR pose regression network called HypLiLoc, which utilizes multimodal features from 3D point clouds and spherical projections, and integrates them in Euclidean and hyperbolic spaces.
I2MVFormer: Large Language Model Generated Multi-View Document Supervision for Zero-Shot Image Classification
Muhammad Ferjad Naeem (ETH Zurich), Federico Tombari
CodeClassificationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: Proposes the I2MVFormer model, which utilizes large language models to generate multi-view text supervision for unsupervised zero-shot image classification.
IDGI: A Framework To Eliminate Explanation Noise From Integrated Gradients
Ruo Yang (Illinois Institute of Technology), Mustafa Bilgic (Illinois Institute of Technology)
CodeExplainability and InterpretabilityConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes the Important Direction Gradient Integration (IDGI) framework, which removes the explanation noise in the Integrated Gradients (IG) method and enhances interpretability.
IFSeg: Image-Free Semantic Segmentation via Vision-Language Model
Sukmin Yun (Mohamed bin Zayed University of Artificial Intelligence), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
CodeSegmentationTransformerSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: Using a pre-trained visual language encoder-decoder model, artificial images are generated from semantic category words to complete the semantic segmentation task without using any task-specific images or annotations.
🎯 What it does: This paper presents Painter, a general visual model that uses image pairs as task prompts to support reasoning for various visual tasks in context.
Imitation Learning As State Matching via Differentiable Physics
Siwei Chen (National University of Singapore), Zhongwen Xu (Sea AI Lab)
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequentialPhysics Related
🎯 What it does: By utilizing a differentiable physics simulator to directly embed the state matching target into gradient descent, single-loop imitation learning is achieved.
🎯 What it does: This paper analyzes the fundamental reasons for the poor performance of deepfake detection models in cross-dataset evaluations and proposes the theory of 'implicit identity leakage'. It then designs an ID-unaware deepfake detection model based on local forgery trace detection.
🎯 What it does: To address the domain generalization problem, this paper proposes an improved test-time adaptation method called ITTA, which dynamically adjusts the model during the testing phase to adapt to the target domain.
Improving Commonsense in Vision-Language Models via Knowledge Graph Riddles
Shuquan Ye (Microsoft), Jing Liao (City University of Hong Kong)
CodeRetrievalTransformerVision Language ModelImageText
🎯 What it does: This paper proposes the DANCE technique, which generates puzzle-like text and image pairs by linearizing conceptual graph knowledge and hiding entities, thereby injecting common sense knowledge into visual-language models during the training phase.
🎯 What it does: Proposes modeling domain generalization as a convex game between domains, enhancing the generalization ability of multi-source domains through hypermodel regularization and sample filtering.
Improving Robustness of Vision Transformers by Reducing Sensitivity To Patch Corruptions
Yong Guo (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
CodeClassificationRecognitionTransformerImage
🎯 What it does: A new training method called RSPC is proposed, specifically designed to enhance the model's robustness against common image perturbations by reducing the sensitivity of the Transformer to patch corruption.
🎯 What it does: This paper studies the selective prediction problem in Visual Question Answering (VQA) and proposes a learning method using peer models (Learning from Your Peers, LYP) to train a selector, enabling the model to self-reject when facing uncertainty in correctness.
Improving Visual Grounding by Encouraging Consistent Gradient-Based Explanations
Ziyan Yang (Rice University), Vicente Ordonez (Rice University)
CodeObject DetectionExplainability and InterpretabilityTransformerVision Language ModelImage
🎯 What it does: This paper proposes a gradient-based interpretable consistency (AMC) loss to directly optimize GradCAM heatmaps in visual language models, aligning them with human-annotated regions to enhance visual localization capabilities.
Improving Zero-Shot Generalization and Robustness of Multi-Modal Models
Yunhao Ge (Google Research), Jiaping Zhao (Google Research)
CodeClassificationPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper addresses zero-shot classification in multimodal image-text models by proposing a confidence estimation method based on prompt and image transformation self-consistency, and enhances labels using the WordNet hierarchy on low-confidence samples to improve prediction accuracy.
Incremental 3D Semantic Scene Graph Prediction From RGB Sequences
Shun-Cheng Wu (Technische Universität München), Federico Tombari (Google)
CodeRecognitionObject DetectionSegmentationComputational EfficiencyGraph Neural NetworkSimultaneous Localization and MappingImageVideoPoint Cloud
🎯 What it does: A real-time incremental 3D semantic scene graph inference framework based on RGB sequences is proposed, which can continuously construct a globally consistent 3D scene graph without relying on depth information.
🎯 What it does: This paper proposes a new multi-task learning optimization method called Aligned-MTL, which eliminates gradient conflicts and dominance issues by aligning the gradient matrices, thereby enhancing the stability and performance of multi-task training.
🎯 What it does: This paper proposes a self-supervised multimodal spatial evaluator (IMSE) to address the distribution discrepancy problem in multimodal image registration, and directly drives the optimization of the registration network through the evaluator's error.
Indiscernible Object Counting in Underwater Scenes
Guolei Sun (ETH Zurich), Luc Van Gool (KU Leuven)
CodeObject DetectionTransformerImage
🎯 What it does: A new task called 'concealed object counting' is proposed, and a large underwater fish concealment counting dataset, IOCfish5K, is constructed. Based on this, a Transformer framework called IOCFormer is proposed, which integrates density and regression branches.
🎯 What it does: This paper proposes a PartNet based on a small number of supporting partial images to infer object part segmentation maps from semantic maps, and enhances the detail quality of semantic image synthesis through a part semantic modulation module.
Instance-Specific and Model-Adaptive Supervision for Semi-Supervised Semantic Segmentation
Zhen Zhao (University of Sydney), Luping Zhou (University of Sydney)
CodeObject DetectionSegmentationImage
🎯 What it does: This paper proposes iMAS, an instance-specific and model-adaptive semi-supervised semantic segmentation method that utilizes a teacher-student model to evaluate the difficulty of each unlabeled sample based on IoU, and dynamically adjusts the weights of strong augmentation and consistency loss according to this difficulty.
🎯 What it does: Directly inferring a complete 3D head mesh from multi-view calibrated images, omitting traditional MVS reconstruction and non-rigid registration steps.
🎯 What it does: Proposes the InstMove module, which improves video instance segmentation, video object segmentation, and multi-object tracking/segmentation through instance-level motion prediction.
🎯 What it does: Reformulate the interactive segmentation task as a pixel-level binary classification problem for each image, and use a Gaussian process classification model.
🎯 What it does: A new large-scale convolutional foundation model, InternImage, is proposed, utilizing deformable convolution to achieve long-range dependencies and adaptive spatial aggregation.
🎯 What it does: A multi-instance learning framework based on causal intervention, IBMIL, is proposed, which can suppress background correlation bias in whole slide image classification and improve bag-level prediction accuracy.
🎯 What it does: This paper proposes a Clean Feature Mixup (CFM) method that mixes clean features in the feature space to introduce a competitive mechanism when generating targeted adversarial samples, thereby enhancing the cross-model transferability of adversarial samples.
🎯 What it does: This study investigates the inverse rendering of the shape, surface reflection, subsurface scattering (SSS), and ambient lighting of translucent objects from flash and non-flash images.
🎯 What it does: This paper proposes an artistic style transfer framework called InST, which is based on reverse learning text descriptions from a single artwork. It utilizes attention-based text reverse and random reverse methods to use the learned style information as a conditional driver for a diffusion model to generate new artistic images, achieving precise transfer of attributes such as semantics, texture, brushstrokes, and color.
Yash Kant (University of Toronto), Igor Gilitschenski (Snap Research)
CodeGenerationPose EstimationMesh
🎯 What it does: An end-to-end differentiable Inverse Neural Skinning (INS) pipeline is proposed, utilizing a Pose-Conditioned Inverse Network (PIN) to capture the nonlinear deformations of clothing and muscles on both sides of differential LBS, achieving surface correspondence during pose re-targeting with only one mesh extraction.
🎯 What it does: This paper proposes a two-stage generative transformer framework IS-GGT for scene graph generation: first, it samples the interaction graph between entities using a generative transformer, and then classifies the predicates of the sampled edges;
🎯 What it does: This paper proposes an Iterative Geometric Encoding Volume (IGEV) framework that generates Geometric Encoding Volumes (GEV) using a lightweight 3D CNN and fuses them with All-Pairs Correlation (APC) to form a Combined Geometric Encoding Volume (CGEV). It then uses ConvGRU to iteratively update the disparity and accelerates convergence by regressing the initial disparity through soft-argmin, extending it to multi-view stereo (MVS) tasks.
🎯 What it does: An iterative next ring detection (INBD) method is proposed for tree ring instance segmentation in cell-level high-resolution microscopic images of tree cross-sections.
Iterative Proposal Refinement for Weakly-Supervised Video Grounding
Meng Cao (Peking University), Daxin Jiang (Microsoft)
CodeKnowledge DistillationRepresentation LearningTransformerVision Language ModelVideo
🎯 What it does: An Iterative Proposal Refinement Network (IRON) is proposed for weakly supervised video grounding, utilizing semantic and conceptual knowledge from a pre-trained VL model for dual distillation, and iteratively updating proposal confidence through label propagation.
🎯 What it does: This paper proposes a knowledge distillation method called itKD for 3D point cloud object detection, aimed at training lightweight detectors.
🎯 What it does: This paper proposes aligning the mutual information (MI) in the gradient space of Neural Radiance Fields (NeRF) so that when the network weights are perturbed, semantically related scene points or regions can produce resonant mutual responses, thereby achieving tasks such as sparse label propagation, instance selection, and editing.
🎯 What it does: This paper proposes and implements a single-stage JAMNet network for recovering high-quality global shutter images from two frames of rolling shutter images.
🎯 What it does: This paper proposes a Token Pruning and Compression (TPS) module to more aggressively compress visual Transformers while retaining the information of pruned tokens.
🎯 What it does: An end-to-end method is proposed to simultaneously perform multi-frame interpolation and deblurring for videos with unknown exposure times.
🎯 What it does: A joint visual localization and tracking framework is proposed, utilizing natural language descriptions to simultaneously achieve target localization and tracking.
🎯 What it does: The k-planes model is proposed, which achieves an explicit representation of radiance fields for 3D static and 4D dynamic scenes by splitting the d-dimensional space into (d choose 2) two-dimensional planes.
KERM: Knowledge Enhanced Reasoning for Vision-and-Language Navigation
Xiangyang Li (Chinese Academy of Sciences), Shuqiang Jiang (Chinese Academy of Sciences)
CodeRetrievalRobotic IntelligenceTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A knowledge-enhanced reasoning model KERM is proposed, which retrieves knowledge facts corresponding to views and integrates visual, historical, and instruction features to improve visual-language navigation.
Knowledge Combination To Learn Rotated Detection Without Rotated Annotation
Tianyu Zhu (Amazon), Anton van den Hengel (Monash University)
CodeObject DetectionDomain AdaptationImage
🎯 What it does: Proposes the KCR framework, which utilizes the joint training of source data's rotation annotations and target data's axis-aligned annotations to achieve rotation object detection without the need for rotation annotations.
🎯 What it does: The Label Information Bottleneck (LIB) method is proposed for label enhancement, restoring the complete label distribution based on logical labels.
🎯 What it does: By synthesizing realistic liver tumors in healthy liver CT scans, AI models can be trained to perform liver tumor segmentation without the need for manual voxel annotation.
LANA: A Language-Capable Navigator for Instruction Following and Generation
Xiaohan Wang (Zhejiang University), Yi Yang (Zhejiang University)
CodeGenerationExplainability and InterpretabilityRobotic IntelligenceTransformerVision Language ModelMultimodality
🎯 What it does: A single model named LANA has been developed, capable of simultaneously executing visual language navigation instructions and generating path descriptions, achieving bidirectional human-machine language interaction.
Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification
Yue Yang (University of Pennsylvania), Mark Yatskar (University of Pennsylvania)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes an interpretable image classification framework called LaBo, which does not require manually defined concepts. It utilizes large language models to generate candidate concepts and aligns them visually through CLIP, forming a concept bottleneck model.
CodeData SynthesisSafty and PrivacyFlow-based ModelVideo
🎯 What it does: A large-capacity, reversible video steganography network is proposed, capable of hiding up to 7 secret videos within a cover video and achieving complete recovery of these secret videos through a single reversible neural network.
Large-Scale Training Data Search for Object Re-Identification
Yue Yao (Australian National University), Liang Zheng (Australian National University)
CodeRecognitionRetrievalImage
🎯 What it does: The paper proposes a Search and Prune (SnP) framework to quickly construct a training set that is close to the target domain distribution and has a controllable size from a large-scale data pool, in order to enhance the performance of object re-identification in the target domain.
LAVENDER: Unifying Video-Language Understanding As Masked Language Modeling
Linjie Li (Microsoft), Lijuan Wang (Microsoft)
CodeGenerationRetrievalTransformerVision Language ModelVideoText
🎯 What it does: LAVENDER is proposed, a unified video-language model that uses Masked Language Modeling (MLM) as a unified interface for all pre-training and downstream tasks, without the need for task-specific heads.
Layout-Based Causal Inference for Object Navigation
Sixian Zhang (Chinese Academy of Sciences), Shuqiang Jiang (Chinese Academy of Sciences)
CodeRobotic IntelligenceReinforcement Learning
🎯 What it does: A layout-based soft total direct effect (L-sTDE) framework is proposed, which adjusts the positive and negative influences of experiences by estimating environmental layout differences in target navigation, enhancing generalization ability in unknown environments.
🎯 What it does: Developed and trained the first 3D deformable facial reflection model with spatially varying BRDF that can learn from low-cost public data.
🎯 What it does: This paper proposes a color difference (CD) metric method based on multi-scale autoregressive normalizing flows—CD-Flow—for evaluating color differences in real photographic images.
Learning a Practical SDR-to-HDRTV Up-Conversion Using New Dataset and Degradation Models
Cheng Guo (Communication University of China), Xiuhua Jiang (Communication University of China)
CodeImage TranslationRestorationTransformerVideo
🎯 What it does: This paper studies the upsampling from SDR video to HDR-WCG television, proposing a new dataset HDRTV4K and an improved HDR→SDR degradation model.
🎯 What it does: This paper proposes an audio-visual source localization method based on False Negative Aware Contrastive Learning (FNAC), which utilizes a unimodal similarity matrix to identify potential false negative samples and suppress their impact through two regularization techniques: False Negatives Suppression (FNS) and True Negatives Enhancement (TNE), thereby enhancing the audio-visual correspondence representation.
Learning Bottleneck Concepts in Image Classification
Bowen Wang (Osaka University), Hajime Nagahara (Osaka University)
CodeClassificationContrastive LearningImage
🎯 What it does: This paper proposes BotCL (Bottleneck Concept Learner), a model that learns interpretable concepts and performs image classification using a slot attention mechanism and self-supervised adversarial learning under the supervision of no concept labels.
🎯 What it does: This paper proposes a decoupled regularization method based on Fast Fourier Transform (FFT) that efficiently decorrelates features in self-supervised visual representation learning, reducing the time complexity to O(n d log d) and significantly lowering computational costs.
🎯 What it does: Proposes a feature refinement head (FR Head) based on contrastive learning to enhance feature representation in skeleton action recognition models, particularly improving the distinction of ambiguous actions.
🎯 What it does: EmotionCLIP is constructed, a framework for visual emotion representation pre-training through unlabeled everyday communication videos and subtitles.
🎯 What it does: The FedPR algorithm is proposed, which learns distributed visual prompts in the approximate null space of global prompts to address the issues of communication costs, data scarcity, and catastrophic forgetting in federated MRI reconstruction.
🎯 What it does: This paper proposes a multi-stage label purification method called Decoupled Meta Label Purifier (DMLP) to enhance the robustness of deep networks in the presence of noisy labels.