Suman Saha (ETH Zurich), Luc Van Gool (ETH Zurich)
CodeSegmentationDomain AdaptationTransformerImage
π― What it does: A framework called EDAPS specifically designed for domain adaptive panoptic segmentation is proposed, achieving end-to-end training on a dataset transitioning from synthetic to real scenes.
π― What it does: A training-free and lightweight fine-tuning framework for person-object interaction detection, ADA-CM, is proposed, utilizing a concept-guided memory module and instance-aware adapter for efficient detection.
π― What it does: This paper proposes a Transformer-based multi-task visual scene understanding framework that utilizes the weights and activations of single-task models for cross-task and cross-time sharing to achieve efficient multi-task inference.
π― What it does: A loss weighting strategy named MinβSNRΞ³ is proposed to adjust the gradient conflicts at different time steps during the training of diffusion models, thereby accelerating convergence and improving generation quality.
π― What it does: The SuperLiDAR network is proposed for end-to-end over-segmentation of LiDAR point clouds, generating superpoints that are uniform in both semantics and geometry.
π― What it does: This paper proposes an efficient video action detection framework EVAD based on ViT, achieving efficient inference through spatiotemporal token dropping centered on key frames and context refinement.
π― What it does: This paper proposes the EfficientViT model, which achieves high-resolution dense prediction through lightweight multi-scale attention.
EGC: Image Generation and Classification via a Diffusion Energy-Based Model
Qiushan Guo (University of Hong Kong), Ping Luo (University of Hong Kong)
CodeClassificationGenerationDiffusion modelImage
π― What it does: A unified energy-based model EGC is proposed, which can perform image classification and generate images through a diffusion reverse process. The forward process estimates the energy of the joint distribution p(x, y), while the backward process computes the gradient of the joint distribution for sample reconstruction and generation.
EgoLoc: Revisiting 3D Object Localization from Egocentric Videos with Visual Queries
Jinjie Mai (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
CodeObject DetectionPose EstimationRetrievalSimultaneous Localization and MappingVideo
π― What it does: An end-to-end visual query 3D localization (VQ3D) pipeline called EgoLoc is proposed to locate the most recent appearance of the queried object in first-person videos and provide a relative 3D displacement vector.
EgoObjects: A Large-Scale Egocentric Dataset for Fine-Grained Object Understanding
Chenchen Zhu (Meta AI), Zhicheng Yan (Meta AI)
CodeObject DetectionFederated LearningVideo
π― What it does: Created a large-scale first-person perspective EgoObjects dataset, containing hundreds of thousands of video frames and target boxes, annotated with category and instance IDs;
π― What it does: We propose EgoVLPv2, a second-generation self-centered video-language pre-training framework that integrates cross-modal attention into the video and text Transformer backbone.
π― What it does: In the visual place recognition task, a training method is proposed that does not require additional annotations, utilizing map partitioning and principal component analysis to automatically construct multi-view classes of the same location, thereby training a global descriptor that is robust to viewpoint changes.
π― What it does: This study proposes ElasticViT, a two-stage neural architecture search method designed for deploying high-accuracy, low-latency visual Transformer (ViT) models on various mobile devices.
π― What it does: This paper proposes a learning-based encoder called ELITE, which quickly and accurately maps visual concepts to text embedding space, enabling customized text-to-image generation.
π― What it does: This paper proposes a customizable and learnable prior deep unrolling framework (CUE) for low-light image enhancement, utilizing Masked Autoencoder pre-trained illumination and noise priors, which are embedded in the proximal operations and regularization terms of the Retinex unrolling steps, thereby achieving a more transparent and interpretable enhancement model.
π― What it does: By directly performing gradient optimization on the noise vector of the diffusion model (DOODL), precise guidance is applied to the final generated images using pre-trained discriminative networks (such as CLIP, FGVC classifiers, and aesthetic scorers), achieving multimodal, personalized, and aesthetic optimization without the need to retrain noise-aware classifiers.
π― What it does: This paper proposes a semi-supervised semantic segmentation framework based on enhanced soft labels (ESL), which combines dynamic soft labels (DSL) with pixel-to-part contrastive learning to fully utilize high-entropy pseudo-label information and improve category boundary recognition capabilities.
Enhancing Adversarial Robustness in Low-Label Regime via Adaptively Weighted Regularization and Knowledge Distillation
Dongyoon Yang (Seoul National University), Yongdai Kim (Seoul National University)
CodeKnowledge DistillationAdversarial AttackImage
π― What it does: In the scenario of limited labeled data, a semi-supervised adversarial training method is proposed that combines adaptive weighted regularization with semi-supervised knowledge distillation;
π― What it does: A fine-tuning strategy based on Sharpness-Aware Minimization (SAM) called FT-SAM is proposed to more effectively eliminate implanted backdoors when only a small number of clean samples are available.
π― What it does: A Sample Adaptive Augmentation (SAA) framework is proposed to identify and apply more diverse augmentations to 'naive' samples in semi-supervised learning;
π― What it does: An Environment-Agnostic Reinforcement Learning (EAR) framework is proposed, which utilizes feature decomposition, reconstruction, and action-based state shift self-supervised objectives to extract environment-agnostic features for visual RL policy learning.
π― What it does: A framework named EICR is proposed, providing a unified solution to the issues of predicate category imbalance and subject-object pair context imbalance in scene graph generation.
eP-ALM: Efficient Perceptual Augmentation of Language Models
Mustafa Shukor (Sorbonne University), Matthieu Cord (Valeo)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityAudio
π― What it does: We propose eP-ALM, a multimodal adaptation method that efficiently integrates large language models with visual/video/audio encoders using only a few linear projection layers and soft prompts, while nearly freezing all parameters.
π― What it does: A flexible quantization neural network (EQ-Net) is proposed, which can generate quantized sub-networks of different bit widths, symmetric/asymmetric, and adjustable granularity through a single training session.
Equivariant Similarity for Vision-Language Foundation Models
Tan Wang (Nanyang Technological University), Lijuan Wang (Microsoft)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageVideoTextMultimodalityBenchmark
π― What it does: This paper proposes and implements Equivariant Similarity Learning (EQSIM) in Vision-Language Models (VLM) and constructs a benchmark focused on visual minor changes (EQBEN), aiming to enhance the model's robustness under fine-grained semantic variations.
π― What it does: A method is proposed for text-to-image diffusion models that allows for the erasure of specified visual concepts using only concept names without any additional data on pre-trained model weights.
Essential Matrix Estimation using Convex Relaxations in Orthogonal Space
Arman Karimian (Boston University), Roberto Tron (Boston University)
CodePose EstimationOptimizationSimultaneous Localization and MappingImage
π― What it does: This paper proposes a new two-view structure from motion (SfM) method that uses a four-dimensional rotation matrix embedding to estimate the essential matrix, achieving a globally optimal solution through semidefinite relaxation and Riemannian gradient descent.
π― What it does: A new binary network training method called ReSTE is proposed, which improves STE to balance estimation error and gradient stability, achieving efficient binary training.
Evaluation and Improvement of Interpretability for Self-Explainable Part-Prototype Networks
Qihan Huang (Zhejiang University), Mingli Song (Zhejiang University)
CodeExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: Proposed quantifiable metrics for evaluating the interpretability of part prototype networks, and based on this, improved ProtoPNet to enhance its prototype consistency and robustness.
π― What it does: An Explicit Motion Disentangling (EMD-Flow) framework is proposed, which explicitly separates global motion learning from local refinement, and achieves efficient optical flow estimation through two lightweight modules: Multi-scale Motion Aggregation (MMA) and Confidence-induced Flow Propagation (CFP).
π― What it does: This paper proposes an end-to-end social navigation architecture that injects common-sense cognition of human-machine interaction into reinforcement learning strategies through two types of 'proximity perception tasks'βRisk Estimation and Proximity Compassβachieving safe and efficient robot navigation in human environments.
Exploring Lightweight Hierarchical Vision Transformers for Efficient Visual Tracking
Ben Kang (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
CodeObject TrackingTransformerVideo
π― What it does: A lightweight hierarchical Vision Transformer tracking framework HiT is proposed, which utilizes a Bridge Module to fuse deep semantic features with shallow detail features, and enhances the relationship between the search and template through dual-image position encoding, constructing an efficient real-time visual tracker.
Exploring Model Transferability through the Lens of Potential Energy
Xiaotong Li (Peking University), Ling-Yu Duan (Peking University)
CodeClassificationDomain AdaptationRepresentation LearningContrastive LearningImagePhysics Related
π― What it does: A physics-driven method based on potential energy descent is proposed to simulate the representation dynamics in the process of transfer learning, thereby improving the transferability assessment of pre-trained models.
π― What it does: This paper proposes an improved two-stage Transformer structure for detecting human-object interactions (HOI), enhancing predicate classification by incorporating box-pair-based positional information and richer image context into the cross-attention mechanism.
Exploring the Benefits of Visual Prompting in Differential Privacy
Yizhe Li (Xi'an Jiaotong University), Xuebin Ren (IBM Research)
CodeClassificationSafty and PrivacyTransformerPrompt EngineeringImage
π― What it does: Utilizing visual prompting to reshape pre-trained models, using them as teacher models, and combining with the PATE framework for differential privacy training to construct a high-accuracy differential privacy image classifier.
π― What it does: Constructed two digital twin datasets, YCB-Real and YCB-Synthetic, systematically introducing five types of defects (noise, holes, texture blur, baked lighting, ambient lighting) into 3D models, and evaluated the impact of these defects on model performance in object detection and instance segmentation tasks, while providing a cost-benefit analysis of artist repair time and model accuracy.
π― What it does: This study constructed the DIVIDE-3k user-generated content video quality database and proposed two no-reference video quality assessment methods, DOVER and DOVER++, based on the separation of aesthetic and technical perspectives.
Extensible and Efficient Proxy for Neural Architecture Search
Yuhong Li (University of Illinois), Deming Chen (University of Illinois)
CodeNeural Architecture SearchImage
π― What it does: This paper proposes a scalable low-cost proxy (Eproxy) and discrete proxy search (DPS) for rapid evaluation of network architectures across different NAS search spaces and multimodal settings.
π― What it does: This paper presents FaceCLIPNeRF, a text-driven 3D avatar NeRF deformation and rendering framework that enables automated editing of facial expressions, emotions, and other attributes while maintaining facial identity and high-quality rendering.
Factorized Inverse Path Tracing for Efficient and Accurate Material-Lighting Estimation
Liwen Wu (University of California San Diego), Ravi Ramamoorthi (University of California San Diego)
CodeComputational EfficiencyImage
π― What it does: This paper proposes a new inverse path tracing methodβFactorized Inverse Path Tracing (FIPT)βfor efficiently and accurately estimating material properties and light distribution in indoor scenes.
π― What it does: This paper proposes a method for automatically identifying subsets of bias conflicts in visual datasets caused by spurious correlations, called FACTS;
π― What it does: This paper proposes a new local descriptor - Fan Beam Binary Differential Projection (FB-BDP) for fine-grained leaf image retrieval;
Fast Full-frame Video Stabilization with Iterative Optimization
Weiyue Zhao (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
CodeRestorationOptimizationOptical FlowVideo
π― What it does: A full-frame video stabilization method based on iterative optimization is proposed, which includes two main modules: motion trajectory smoothing and full-frame outpainting.
π― What it does: A runtime optimization method for neural scene flow estimation based on distance transformation is implemented, which achieves real-time performance on dense point clouds without the need for training.
π― What it does: A front-and-back projection joint BEV transformation module FB-BEV is proposed to generate denser and higher-quality bird's-eye view features for multi-camera 3D detection.
FCCNs: Fully Complex-valued Convolutional Networks using Complex-valued Color Model and Loss Function
Saurabh Yadav (Indraprastha Institute of Information Technology Delhi), Koteswar Rao Jerripothula (Indraprastha Institute of Information Technology Delhi)
π― What it does: A Full Complex-valued Convolutional Network (FCCN) is proposed, achieving a complete flow of complex numerical information from input to output.
π― What it does: A post-processing method is proposed to eliminate artifacts in StyleGAN synthesized images by identifying and scaling 'cancer' features, enhancing image quality without sacrificing diversity.
π― What it does: This paper studies partial model personalization of Vision Transformers (ViT) in federated learning. It first identifies that the self-attention layers and classification heads are the most sensitive layers through empirical research, and then proposes the FedPerfix method, which inserts a Prefix plugin into the global self-attention layer and uses parallel adapters for stable initialization, thereby achieving a mix of local and global attention learning.
π― What it does: An end-to-end deep learning framework called FineRecon is proposed, which directly predicts the scene TSDF from multi-view images with known poses, achieving high-fidelity 3D reconstruction.
FLIP: Cross-domain Face Anti-spoofing with Language Guidance
Koushik Srivatsan (Mohamed Bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed Bin Zayed University of Artificial Intelligence)
CodeRecognitionDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: This paper studies a cross-domain face spoofing detection method based on a vision-language pre-training model.
Yuning Cui (Technical University of Munich), Alois Knoll (Technical University of Munich)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: An efficient image restoration network named FocalNet is proposed, primarily targeting tasks such as dehazing, snow removal, and deblurring.
π― What it does: Proposes a dual teacher-student framework and introduces a bidirectional learning strategy, separating the two capabilities of learning and adaptation, enhancing the performance of unsupervised domain adaptive semantic segmentation.
π― What it does: A self-supervised pre-training framework called Forecast-MAE based on Masked Autoencoder is proposed for traffic motion prediction tasks.
π― What it does: This paper proposes a cross-attention fusion framework based on foreground and text lines to achieve geometric distortion removal of document images and improve readability.
Foreground-Background Separation through Concept Distillation from Generative Image Foundation Models
Mischa Dombrowski (Friedrich-Alexander-UniversitΓ€t Erlangen-NΓΌrnberg), Bernhard Kainz (Imperial College London)
CodeSegmentationGenerationDiffusion modelImageBiomedical Data
π― What it does: Automatically generate foreground-background segmentation masks through a pre-trained latent diffusion model and text prompts, achieving unsupervised segmentation model training.
π― What it does: This paper proposes an unsupervised curve structure segmentation method called FreeCOS, which extracts robust features from fractals and unlabeled images through self-supervised learning to achieve segmentation of curved objects such as blood vessels and cracks.
π― What it does: This paper proposes two unified knowledge distillation methods: Normalized KD (NKD) and Universal Self-Knowledge Distillation (USKD), which are used for teacher-assisted and teacher-free training scenarios, respectively.
π― What it does: A large-scale high-quality real rain dataset, LHP-Rain, has been constructed, and the RLRTR video de-raining and SCD-Former single image de-raining models have been proposed.
GACE: Geometry Aware Confidence Enhancement for Black-Box 3D Object Detectors on LiDAR-Data
David Schinagl (Graz University of Technology), Horst Bischof (Graz University of Technology)
CodeObject DetectionAutonomous DrivingPoint Cloud
π― What it does: Proposes the GACE method, which performs geometry information-driven post-processing on the confidence estimation of black-box LiDAR 3D detectors to enhance detection performance.
GAFlow: Incorporating Gaussian Attention into Optical Flow
Ao Luo (Megvii Technology), Shuaicheng Liu (University of Electronic Science and Technology of China)
CodeTransformerOptical FlowImageVideo
π― What it does: This paper proposes the GAFlow framework, which introduces a learnable Gaussian attention module in optical flow estimation, used to enhance the local discriminability of feature representations (Gaussian-Constrained Layer, GCL) and the motion affinity during the matching process (Gaussian-Guided Attention Module, GGAM).
π― What it does: This paper proposes the GaPro method, which utilizes axis-aligned 3D bounding box supervision to generate pseudo-instance masks through Gaussian processes, and uses these to train a 3D point cloud instance segmentation network.
GasMono: Geometry-Aided Self-Supervised Monocular Depth Estimation for Indoor Scenes
Chaoqiang Zhao (East China University of Science and Technology), Stefano Mattoccia (University of Bologna)
CodeDepth EstimationKnowledge DistillationTransformerSimultaneous Localization and MappingImage
π― What it does: The GasMono framework is proposed to address the issues of large rotations and low texture in indoor scenes through geometric-assisted self-supervised monocular depth estimation.
π― What it does: This paper proposes a ground-embedded module GEDepth, which decouples camera parameters and image features to enhance the generalization ability of monocular depth estimation.
π― What it does: This paper proposes a framework called Visual Concept Translator (VCT) for general image-to-image translation tasks given only a reference image, capable of transferring visual concepts from the reference image while preserving the content of the source image.
General Planar Motion from a Pair of 3D Correspondences
Juan Carlos Dibene (Stevens Institute of Technology), Enrique Dunn (Stevens Institute of Technology)
CodePose EstimationAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
π― What it does: A geometric closed-form solver based on two 3D-3D corresponding points is proposed to estimate the relative pose (5 degrees of freedom) of the camera and the motion plane under unknown motion planes.
π― What it does: This paper proposes a dual meta-learning framework MEDIC for open set domain generalization, which jointly learns model parameters through domain-level and class-level gradient matching, and constructs decision boundaries for each known class using multiple binary classifiers.
Tong Wei (Czech Technical University in Prague), Daniel Barath (ETH Zurich)
CodeOptimizationReinforcement LearningPoint Cloud
π― What it does: A differentiable RANSAC framework, β-RANSAC, is proposed and implemented, capable of learning matching confidence, sampling distribution, and minimum solvers, achieving end-to-end training from feature matching to model estimation.
π― What it does: A general photometric adaptive Channel Selective Normalization (CSNorm) module is proposed, which can effectively enhance, color-correct, and beautify images under unknown photometric conditions after being trained with a single photometric condition.
π― What it does: A no-pool domain adaptive prompt (DAP) framework is proposed, capable of generating prompts instantaneously at the instance level to adjust the frozen ViT backbone, enabling replay-free continual learning.
π― What it does: This study investigates the recovery of high-resolution, large-batch, and complex image samples through gradient inversion in federated learning, proposing a CI-Net generator based on over-parameterized convolutional networks.
Generative Novel View Synthesis with 3D-Aware Diffusion Models
Eric R. Chan (Stanford University), Gordon Wetzstein (NVIDIA)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: A new 3D-aware view synthesis framework based on diffusion models is proposed, capable of generating diverse and geometrically consistent novel views from a single or a few input images, and supports autoregressive long sequence generation.
Generative Prompt Model for Weakly Supervised Object Localization
Yuzhong Zhao (University of Chinese Academy of Sciences), Fang Wan (University of Chinese Academy of Sciences)
CodeObject DetectionTransformerVision Language ModelDiffusion modelImage
π― What it does: A generative prompt model called GenPromp is proposed, redefining the weakly supervised object localization problem as a conditional image denoising task. By learning category representative prompt embeddings and fusing them with CLIP discriminative embeddings, high-quality localization maps are generated using multi-scale cross-attention.
π― What it does: A learning-based surface reconstruction framework called GeoUDF is proposed, which can directly reconstruct closed or non-closed 3D surfaces from sparse point clouds and generate high-quality triangular meshes.
π― What it does: This paper proposes a generative model named GEPSAN, which predicts multiple possible next steps from procedural videos (using cooking as an example) and achieves zero-shot transfer under video input.
π― What it does: Generalize the category vector to a linear subspace and use subspace projection in the final fully connected layer to compute logits, constructing the Grassmann Class Representation (GCR);
π― What it does: A gradient inversion attack method based on Generative Adversarial Networks (GAN) is proposed (GIFD), which recovers the private data corresponding to the uploaded gradients in federated learning by searching the feature domain of the generator layer by layer.
GLA-GCN: Global-local Adaptive Graph Convolutional Network for 3D Human Pose Estimation from Monocular Video
Bruce X.B. Yu (Hong Kong Polytechnic University), Chang Wen Chen (Hong Kong Polytechnic University)
CodePose EstimationGraph Neural NetworkVideo
π― What it does: A global-local adaptive graph convolutional network (GLA-GCN) is proposed for 3D human pose estimation in monocular videos, capturing spatiotemporal structures in the global layer and performing fine regression for each joint in the local layer.
π― What it does: Proposes the SuperGlobal system, which utilizes improved global features to complete image retrieval and re-ranking, completely independent of local features;
Global Perception Based Autoregressive Neural Processes
Jinyang Tai (Shanghai University)
CodeGenerationData SynthesisMeta LearningRecurrent Neural NetworkTransformerAuto EncoderImageTime Series
π― What it does: A self-autoregressive neural process framework AENPs and CAENPs is proposed, improving the latent distribution and deterministic paths of NPs, allowing the model to better capture the global and local relationships of contextual sample points.
Gloss-Free Sign Language Translation: Improving from Visual-Language Pretraining
Benjia Zhou (MUST), Du Zhang (MUST)
CodeRecognitionGenerationTransformerVision Language ModelContrastive LearningVideoText
π― What it does: This paper proposes a gloss-free supervised sign language translation framework GFSLT-VLP, which utilizes visual-language pre-training (VLP) and masked self-supervised learning to enhance the language-guided representation of the visual encoder, thereby achieving end-to-end translation from sign language to text.
π― What it does: This paper proposes the GlueGen framework, which utilizes GlueNet to achieve seamless alignment of different conditional encoders (such as multilingual and audio) with existing stable diffusion models, enabling X-to-image generation.
Going Beyond Nouns With Vision & Language Models Using Synthetic Data
Paola Cascante-Bonilla (Rice University), Leonid Karlinsky (IBM Research)
CodeData SynthesisDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: By generating millions of synthetic image-text pairs (SyViC dataset) and combining techniques such as parameter-efficient fine-tuning (LoRA), domain-adaptive style transfer, and long text chunking, we fine-tune existing large-scale vision-language models (such as CLIP, CyCLIP) to enhance their understanding of visual language concepts beyond nouns (attributes, relationships, states) and their ability for compositional reasoning.
π― What it does: In visual classification tasks, a multi-head lightweight classifier is introduced, and Gramian attention is used to enhance class labels, improving the model's expressive capability.
π― What it does: The Bi-Level Noisy Correspondence (BNC) problem in graph matching is proposed, and the COMMON method is introduced to achieve robust matching through contrastive learning and momentum distillation.
π― What it does: A graph-based unsupervised domain adaptation method called GraphEcho is proposed for structural segmentation of cardiac ultrasound videos.
GridMM: Grid Memory Map for Vision-and-Language Navigation
Zihan Wang (Chinese Academy of Sciences), Shuqiang Jiang (Chinese Academy of Sciences)
CodeTransformerReinforcement LearningVision Language ModelMultimodality
π― What it does: A dynamic growth grid memory map (GridMM) is proposed to structurally record historical environments in visual and language navigation tasks, and an instruction-related aggregation method is designed to capture fine-grained visual cues.
π― What it does: The GridPull method is proposed, which achieves efficient surface reconstruction of large-scale point clouds by directly optimizing the distance field on a discrete grid, avoiding the use of neural networks.
Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation
Yibo Cui (Defense Innovation Institute), Erwei Yin (Defense Innovation Institute)
CodeTransformerVision Language ModelMultimodality
π― What it does: By constructing a high-quality entity-landmark alignment dataset GEL-R2R, and based on this, conducting three entity-landmark level adaptive pre-training tasks (entity phrase prediction, landmark box prediction, semantic alignment) for the VLN pre-training model, the fine-grained cross-modal alignment ability of visual and language navigation is enhanced.
π― What it does: Proposes to locate interactive areas in 3D point clouds from 2D interactive images, addressing the problem of operability recognition for 3D objects.
π― What it does: The Human3.6M 3D WholeBody (H3WB) dataset is proposed, and based on this, three benchmark tasks for 3D full-body pose estimation are defined; multi-stage baseline results are also provided.
π― What it does: This study proposes a continuous-time model based on hierarchical dynamic graph ODE for predicting future 3D human poses from multi-person 2D skeleton sequences.
Heterogeneous Forgetting Compensation for Class-Incremental Learning
Jiahua Dong (State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences), Gan Sun (State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences)
π― What it does: The Heterogeneous Forgetting Compensation (HFC) model is proposed in class-incremental learning to address the different forgetting rates of old classes at both the representation and gradient levels.
π― What it does: A visual information hiding method based on Type-I adversarial attacks (AVIH) is designed to completely obscure the visual content of images without altering the service model, while maintaining the original functionalities (such as face recognition and classification).