ICCV 2025 Papers — Page 8
IEEE/CVF International Conference on Computer Vision · 2701 papers
EEGMirror: Leveraging EEG Data in the Wild via Montage-Agnostic Self-Supervision for EEG to Video Decoding
Xuan-Hao Liu (Shanghai Jiao Tong University), Wei-Long Zheng (Shanghai Jiao Tong University)
RecognitionGenerationTransformerDiffusion modelContrastive LearningVideoMultimodality
🎯 What it does: An end-to-end framework named EEGMirror is proposed to reconstruct dynamic videos from EEG signals.
Effective Training Data Synthesis for Improving MLLM Chart Understanding
Yuwei Yang (Australian National University), Liang Zheng (Ohio State University)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: This paper constructs a high-quality chart training set ECD through a five-step modular process (single chart generation, composite subchart generation, visual diversification, quality filtering, QA generation), which includes over 10,000 chart images, over 300,000 question-answer pairs, covering 29 types of charts, 25 themes, and over 250 subchart combinations.
Efficient Adaptation of Pre-trained Vision Transformer underpinned by Approximately Orthogonal Fine-Tuning Strategy
Yiting Yang (Xi'an University of Architecture and Technology), Hengtao Shen (TongJi University)
ClassificationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes an AOFT strategy that generates an approximate orthogonal lower/upper projection matrix using only one learnable vector, improving the parameter-efficient fine-tuning of ViT.
Efficient Autoregressive Shape Generation via Octree-Based Adaptive Tokenization
Kangle Deng (Carnegie Mellon University), Tinghui Zhou (Roblox)
GenerationData SynthesisTransformerLarge Language ModelPoint CloudMesh
🎯 What it does: This paper proposes a shape tokenization method based on adaptive octree segmentation (OAT) and the corresponding autoregressive generative model OctreeGPT, achieving efficient tokenization of 3D shapes and text-to-3D generation.
Efficient Concertormer for Image Deblurring and Beyond
Pin-Hung Kuo (National Taiwan University), Ming-Hsuan Yang (University of California)
RestorationTransformerImage
🎯 What it does: This paper proposes an efficient Transformer structure—Concertormer—for single-frame image deblurring, raindrop removal, JPEG compression artifact restoration, and other recovery tasks.
Efficient Event Camera Data Pretraining with Adaptive Prompt Fusion
Quanmin Liang (Sun Yat-Sen University), Yonghong Tian (Peking University)
ClassificationObject DetectionSegmentationTransformerSupervised Fine-TuningPrompt EngineeringOptical FlowImageVideo
🎯 What it does: A lightweight pre-training scheme for event camera data, STP, is proposed, which aligns event data with image models through prompt fusion using image pre-training models.
Efficient Fine-Tuning of Large Models via Nested Low-Rank Adaptation
Lujun Li (Hong Kong University of Science and Technology), Yike Guo
GenerationOptimizationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper presents NoRA, a nested low-rank adaptation structure that significantly reduces trainable parameters and achieves efficient fine-tuning by freezing the outer layer of LoRA and employing serial training of LoRA within it.
Efficient Input-level Backdoor Defense on Text-to-Image Synthesis via Neuron Activation Variation
Shengfang Zhai (Peking University), Jiaheng Zhang (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: A novel input-level backdoor defense framework called NaviT2I is proposed, which can detect and prevent malicious triggering inputs in text-to-image diffusion models based on neuron activation variations.
Efficient Multi-Person Motion Prediction by Lightweight Spatial and Temporal Interactions
Yuanhong Zheng (Shandong University), Jian Sun (Xi'an Jiaotong University)
Pose EstimationComputational EfficiencyTransformerVideo
🎯 What it does: A lightweight multi-person human motion prediction model EMPMP is proposed, which utilizes a dual-branch (local-global) structure to extract local and global spatiotemporal features, and achieves efficient spatial-temporal feature learning through cross-layer interaction.
Efficient Spiking Point Mamba for Point Cloud Analysis
Peixi Wu (University of Science and Technology of China), Xiaoyan Sun (University of Science and Technology of China)
RecognitionSegmentationComputational EfficiencySpiking Neural NetworkPoint Cloud
🎯 What it does: Designed and implemented Spiking Point Mamba (SPM), a 3D point cloud analysis SNN based on the Mamba state space model;
Efficient Track Anything
Yunyang Xiong (Meta AI Research), Vikas Chandra (Meta AI Research)
Object TrackingSegmentationComputational EfficiencyTransformerVideo
🎯 What it does: EfficientTAMs is proposed, a video object segmentation and Track Anything model that utilizes a lightweight ViT encoder and efficient memory cross-attention, significantly reducing model size and inference latency.
Efficient Unsupervised Shortcut Learning Detection and Mitigation in Transformers
Lukas Kuhn (Goethe University), Gemma Roig (Goethe University)
ClassificationAnomaly DetectionComputational EfficiencyTransformerLarge Language ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A framework for unsupervised shortcut learning detection and mitigation based on Transformer is proposed, which can automatically discover the model's excessive reliance on irrelevant features and eliminate this reliance by cropping image patches while retaining the discriminative ability of core features.
Efficient Visual Place Recognition Through Multimodal Semantic Knowledge Integration
Sitao Zhang (Pennsylvania State University), Yelin Kim (Amazon)
RecognitionRetrievalComputational EfficiencyKnowledge DistillationTransformerVision Language ModelImageMultimodality
🎯 What it does: By introducing a pre-trained vision-language model during the training phase, SemVPR simultaneously learns local visual and semantic descriptors, achieving semantic-aware global aggregation during inference to generate extremely compact global features.
EfficientMT: Efficient Temporal Adaptation for Motion Transfer in Text-to-Video Diffusion Models
Yufei Cai (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
GenerationData SynthesisComputational EfficiencyDiffusion modelVideoText
🎯 What it does: We propose EfficientMT, a universal motion transfer framework that converts pre-trained text-to-video diffusion models without the need for optimization during testing.
EFTViT: Efficient Federated Training of Vision Transformers with Masked Images on Resource-Constrained Clients
Meihan Wu (National University of Defense Technology), Xiaodong Wang (National University of Defense Technology)
Federated LearningComputational EfficiencyTransformerImage
🎯 What it does: This paper proposes EFTViT, a hierarchical federated learning framework that efficiently trains Vision Transformer (ViT) models on resource-constrained clients using masked images.
EgoAdapt: Adaptive Multisensory Distillation and Policy Learning for Efficient Egocentric Perception
Sanjoy Chowdhury (University of Maryland), Ruohan Gao (University of Maryland)
RecognitionComputational EfficiencyKnowledge DistillationRecurrent Neural NetworkReinforcement LearningContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes the EGOADAPT framework, which combines cross-modal distillation with adaptive strategy learning to achieve efficient inference for multi-perception egocentric tasks (action recognition, active speaker localization, behavior prediction).
EgoAgent: A Joint Predictive Agent Model in Egocentric Worlds
Lu Chen (Zhejiang University), Sida Peng (Zhejiang University)
Pose EstimationRetrievalRobotic IntelligenceTransformerAgentic AIContrastive LearningImageVideo
🎯 What it does: This paper proposes EgoAgent, a unified model that can simultaneously learn egocentric visual representations, predict future world states, and generate 3D human actions.
Egocentric Action-aware Inertial Localization in Point Clouds with Vision-Language Guidance
Mingfang Zhang (University of Tokyo), Yoichi Sato (University of Tokyo)
RecognitionPose EstimationRobotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkVision-Language-Action ModelContrastive LearningMultimodalityPoint Cloud
🎯 What it does: This paper proposes an inertial positioning framework named EAIL, which utilizes self-perspective motion cues captured by a head-mounted IMU to locate target individuals in a known 3D point cloud environment, with action recognition as an ancillary function.
EgoM2P: Egocentric Multimodal Multitask Pretraining
Gen Li, Siyu Tang (ETH Zurich)
Object TrackingGenerationData SynthesisDepth EstimationTransformerContrastive LearningVideoMultimodality
🎯 What it does: A large-scale forward-looking multimodal multitask pre-training model, EgoM2P, is proposed, which unifies and integrates four modalities: RGB, depth, gaze, and camera trajectory, achieving one-time inference on tasks such as camera tracking, gaze prediction, monocular depth estimation, and conditional video synthesis.
EgoMusic-driven Human Dance Motion Estimation with Skeleton Mamba
Quang Nguyen (FPT Software AI Center), Anh Nguyen
GenerationPose EstimationTransformerDiffusion modelVideoMultimodality
🎯 What it does: A new task is proposed - jointly predicting dance movements from first-person perspective videos and music, and a corresponding dataset EgoAIST++ is constructed.
egoPPG: Heart Rate Estimation from Eye-Tracking Cameras in Egocentric Systems to Benefit Downstream Vision Tasks
Björn Braun, Christian Holz (ETH Zurich)
Convolutional Neural NetworkVideoMultimodalityElectrocardiogram
🎯 What it does: This paper proposes the PulseFormer model to estimate heart rate in real-time by combining an egocentric eye-tracking camera with an IMU sensor, using heart rate as supplementary information to enhance the performance of egocentric visual tasks.
EMatch: A Unified Framework for Event-based Optical Flow and Stereo Matching
Pengjie Zhang (Beijing Institute of Technology), Hua Huang (Beijing Institute of Technology)
Depth EstimationRecurrent Neural NetworkOptical FlowImageVideo
🎯 What it does: A unified framework EMatch is proposed, which can simultaneously estimate optical flow and stereo matching in the same model;
Embodied Image Captioning: Self-supervised Learning Agents for Spatially Coherent Image Descriptions
Tommaso Galliena (Istituto Italiano di Tecnologia), Lorenzo Natale (Istituto Italiano di Tecnologia)
Object DetectionSegmentationGenerationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningContrastive LearningImageTextPoint Cloud
🎯 What it does: A three-stage self-supervised framework is proposed: first, the robot autonomously navigates and collects multi-view images along with their incomplete or inconsistent captions; second, a large language model is utilized to learn to generate consistent pseudo-captions based on caption frequency and context; finally, the pseudo-captions are used to fine-tune an existing image captioning model with contrastive learning (triplet loss) to enhance multi-view consistency.
Embodied Navigation with Auxiliary Task of Action Description Prediction
Haru Kondoh (Institute of Science Tokyo RIKEN AIP), Asako Kanezaki (Institute of Science Tokyo RIKEN AIP)
Explainability and InterpretabilityKnowledge DistillationRobotic IntelligenceTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A framework called DescRL is proposed for generating action descriptions simultaneously in reinforcement learning navigation, achieving an interpretable navigation system.
Embodied Representation Alignment with Mirror Neurons
Wentao Zhu (Peking University), Yizhou Wang (Peking University)
Representation LearningRobotic IntelligenceContrastive LearningVideoText
🎯 What it does: This paper proposes a unified framework inspired by mirror neurons, jointly training two types of models for action understanding and embodied execution, achieving representation alignment through contrastive learning in a shared latent space, thereby enhancing the performance and generalization ability of both tasks.
Embodied VideoAgent: Persistent Memory from Egocentric Videos and Embodied Sensors Enables Dynamic Scene Understanding
Yue Fan (State Key Laboratory of General Artificial Intelligence), Qing Li (State Key Laboratory of General Artificial Intelligence)
Object DetectionObject TrackingRobotic IntelligenceAgentic AIVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes the Embodied VideoAgent, which constructs a persistent 3D object memory and supports active interaction by integrating first-person video with depth, pose, and other body-sensing sensors.
EmbodiedOcc: Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding
Yuqi Wu (Tsinghua University), Jiwen Lu (Tsinghua University)
SegmentationDepth EstimationRobotic IntelligenceGaussian SplattingPoint Cloud
🎯 What it does: Proposes the EmbodiedOcc framework, achieving online 3D volume occupancy prediction based on monocular RGB.
EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device
Gunjan Chhablani (Georgia Tech), Zsolt Kira (Georgia Tech)
Domain AdaptationRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningGaussian SplattingPoint CloudMesh
🎯 What it does: Quickly reconstruct 3D meshes from scenes captured by mobile phones using 3D Gaussian Splatting, and fine-tune navigation strategies in Habitat-Sim to achieve personalized real-to-sim-to-real image target navigation;
EMD: Explicit Motion Modeling for High-Quality Street Gaussian Splatting
Xiaobao Wei (Peking University), Shanghang Zhang (NIO)
Data SynthesisAutonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: For dynamic street scenes in autonomous driving scenarios, an Explicit Motion Decomposition (EMD) module is proposed, which can explicitly model dynamic objects with different motion speeds, significantly improving the reconstruction and view synthesis quality of 3D/4D Gaussian splatting.
EmotiCrafter: Text-to-Emotional-Image Generation based on Valence-Arousal Model
Shengqi Dang (Tongji University), Nan Cao (Tongji University)
GenerationTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes a Continuous Emotion Image Generation task (C‑EICG) and designs the EmotiCrafter model, which integrates Valence-Arousal (V‑A) values with free-text prompts to generate emotionally rich images.
EMoTive: Event-guided Trajectory Modeling for 3D Motion Estimation
Zengyu Wan (University of Science and Technology of China), Zhengjun Zha (University of Science and Technology of China)
Pose EstimationDepth EstimationAutonomous DrivingRecurrent Neural NetworkOptical FlowImageVideo
🎯 What it does: The EMoTive framework is proposed, which models pixel spatiotemporal trajectories using event-driven non-uniform parameter curves to achieve 3D action estimation.
Emulating Self-attention with Convolution for Efficient Image Super-Resolution
Dongheon Lee (University of Seoul), Youngmin Ro (University of Seoul)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a lightweight image super-resolution network called ESC, which replaces most self-attention with a convolutional attention module (ConvAttn) and utilizes Flash Attention to achieve large-window self-attention, significantly reducing computational and memory costs.
End-to-End Driving with Online Trajectory Evaluation via BEV World Model
Yingyan Li (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
Autonomous DrivingTransformerWorld ModelMultimodality
🎯 What it does: The WoTE framework is proposed, which enhances the safety of end-to-end driving by constructing a world model in the BEV space for online trajectory evaluation.
End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation
Liwei Wang (Wuhan Institute of Technology), Huabing Zhou (Wuhan University)
RecognitionObject DetectionTransformerPrompt EngineeringVideo
🎯 What it does: An end-to-end Associative Reasoning Network (ARN) is proposed, which simultaneously performs entity detection and relationship (triplet) prediction in videos, utilizing CLIP's semantic priors, a Predicate Association Parsing module, and a Hierarchical Attention mechanism to achieve associative reasoning of visual relationships.
End-to-End Multi-Modal Diffusion Mamba
Chunhao Lu (China University of Petroleum), Jake Luo (University of Wisconsin)
GenerationData SynthesisDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: An end-to-end multimodal diffusion MAM model (MDM) is proposed, which achieves high-resolution image and long text synchronous generation through VAE unified encoding-decoding and Mamba multi-step selection diffusion.
Engage for All: Making Ordinary Image Descriptions Appealing Again!
Yuyan Chen, Qingpei Guo (Ant Group)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the task of generating appealing descriptions for ordinary images, constructs a large-scale AppealImage dataset, and introduces the CharmNet framework to generate attractive descriptions.
Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation
Haihao Zhang (Beijing Institute of Technology), Xiangyang Ji (Tsinghua University)
Depth EstimationAutonomous DrivingKnowledge DistillationImage
🎯 What it does: Utilizing powerful intensity image information, a cross-sensor knowledge distillation framework is designed to achieve dense disparity estimation using only event streams.
Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions
Dong Li (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A quaternion-based spatial-spectral interaction network, QuatPanNet, is proposed for high-resolution fusion (pansharpening) of multispectral images.
Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling
Zenghao Niu (Shenzhen University), Linlin Shen (Shenzhen University)
Adversarial AttackConvolutional Neural NetworkTransformerLarge Language ModelImageMultimodality
🎯 What it does: Proposes a Gradient-Guided Sampling (GGS) internal iterative sampling strategy, combining MI-FGSM and other iterative attack methods to enhance the balance of black-box transferable attacks.
Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features
Chancharik Mitra (Carnegie Mellon University), Roei Herzig (MIT-IBM Watson AI Lab)
ClassificationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Without gradient fine-tuning, sparse features (Sparse Attention Vectors, SAV) are extracted from the attention heads of large multimodal generative models using a small number of samples (≈20 per label) and are directly applied to visual-language classification tasks.
Enhancing Image Restoration Transformer via Adaptive Translation Equivariance
JiaKui Hu (Peking University), Yanye Lu (Peking University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: An adaptive Transformer based on translation equivariance (TEAFormer) is proposed for various image restoration tasks, including super-resolution, deblurring, denoising, and multi-task recovery.
Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction
Mang Cao (Xi'an Jiaotong University), Le Wang (Xi'an Jiaotong University)
SegmentationDepth EstimationTransformerImage
🎯 What it does: This paper proposes a bidirectional interactive multi-task model called BIM based on Mamba for dense prediction tasks.
Enhancing Numerical Prediction of MLLMs with Soft Labeling
Pei Wang (Amazon), Ashwin Swaminathan (Amazon)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: A soft labeling method is proposed to improve the loss function of multimodal large language models in numerical prediction tasks.
Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning
Jun Li (Harbin Institute of Technology), Bin Chen (Harbin Institute of Technology)
RetrievalTransformerContrastive LearningVideoText
🎯 What it does: This paper proposes a video retrieval framework called HLFormer based on dual spaces (Euclidean + hyperbolic) to address the problem of partial relevant video retrieval (PRVR).
Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection
Xuehan Chen (Xi'an Jiaotong-Liverpool University), Hengyan Liu (Xi'an Jiaotong-Liverpool University)
Object DetectionSegmentationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: By utilizing BLIP to generate multimodal prompts and employing an Adaptive Refinement Module (ARM) to improve the initial prompts, the segmentation performance for concealed targets is enhanced;
Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment
Ying Ba (Renmin University of China), Ji-Rong Wen (Renmin University of China)
GenerationData SynthesisDiffusion modelContrastive LearningImageText
🎯 What it does: In response to the shortcomings of existing reward models in high-quality image evaluation, this paper proposes a new evaluation framework based on Image Containing Text (ICT) scoring, and further trains a human preference (HP) model using only the image modality to better measure human aesthetic preferences, thereby improving the quality of text-to-image generation models.
Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation
Zhenhua Ning (Harbin Institute of Technology), Li Jiang (Chinese University of Hong Kong)
RecognitionSegmentationTransformerLarge Language ModelMultimodalityPoint Cloud
🎯 What it does: This paper proposes a reasoning-based segmentation framework R²S to enhance the performance of 3D multimodal large language models in spatial reasoning.
Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching
Zhankai Li (Central South University), Song Guo (Hong Kong University of Science and Technology)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a new attack framework called ITDS, aimed at significantly enhancing the transferability of targeted adversarial examples.
Enhancing Transformers Through Conditioned Embedded Tokens
Hemanth Saratchandran (Australian Institute for Machine Learning), Simon Lucey
ClassificationObject DetectionSegmentationTransformerImageText
🎯 What it does: Conditioning the attention blocks of the Transformer by adding correction terms to the embedded tokens to reduce their condition number, thereby improving training stability and effectiveness.
Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention
Shiwei Zhang (Xi'an Jiaotong University), Wei Ke (Pengcheng Laboratory)
Object DetectionTransformerVision Language ModelImage
🎯 What it does: This study investigates a general strategy to enhance the accuracy of zero-shot object counting through text-guided local ranking and digit-activated global attention.
Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need
Yongchuan Cui (Aerospace Information Research Institute, Chinese Academy of Sciences), Hui Zhang (Beihang University)
RestorationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes UniPAN, a unified distribution preprocessing strategy that uses inverse transform sampling to map pixels from different satellites to the same target distribution, thereby enhancing the generalization ability of deep learning models across different sensors.
Enrich and Detect: Video Temporal Grounding with Multimodal LLMs
Shraman Pramanick (Meta), Triantafyllos Afouras
RecognitionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: A two-stage video temporal localization method ED-VTG based on a multimodal large language model (LLM) has been developed, which first enriches the input query using LLM and then accurately locates the video through a lightweight interval decoder.
Ensemble Foreground Management for Unsupervised Object Discovery
Ziling Wu (University of Nottingham), Praminda Caleb-Solly (University of Nottingham)
Object DetectionSegmentationAnomaly DetectionKnowledge DistillationTransformerImage
🎯 What it does: Proposes two methods, UnionCut and UnionSeg, to reliably detect image foreground sets in unsupervised object discovery, thereby distinguishing between foreground and background and deciding when to stop discovery.
Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment
RenYe Yan, Ru Huang (Peking University)
GenerationOptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: The AdaEnt model is proposed, which utilizes entropy regularization to dynamically balance exploration and convergence during the fine-tuning of diffusion models in reinforcement learning. It achieves adaptive entropy weighting and dynamic denoising step clipping through a perception model, thereby reducing diversity loss caused by reward mining.
Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation
Shaobo Zhang (Northwest University), Jinye Peng (Northwest University)
Pose EstimationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposes the EA6D framework, which uses a diffusion model to generate environment-independent object representations, and then predicts 6D position and orientation through a pose decoder.
Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation
Chen Gao (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
Depth EstimationImage
🎯 What it does: A light field unsupervised depth estimation network based on disparity-consistent attention aggregation is proposed.
Epona: Autoregressive Diffusion World Model for Autonomous Driving
Kaiwen Zhang, Wei Yin
GenerationAutonomous DrivingTransformerDiffusion modelRectified FlowWorld ModelImageVideoMultimodality
🎯 What it does: Epona is proposed, a self-regressive diffusion-based world model that can generate minute-level future driving scenes and trajectories in a high-resolution and controllable manner given historical driving context.
EquiCaps: Predictor-Free Pose-Aware Pre-Trained Capsule Networks
Athinoulla Konstantinou (University of Aberdeen), Aiden Durrant (University of Lincoln)
Pose EstimationContrastive LearningPoint Cloud
🎯 What it does: This paper proposes EquiCaps, a self-supervised learning framework for pose that leverages the inherent pose awareness of capsule networks without the need for additional predictors.
Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection
Shizhen Zhao, Xiaojuan Qi (University of Hong Kong)
Anomaly DetectionTransformerMixture of ExpertsImage
🎯 What it does: This paper proposes a hybrid expert architecture based on Mixture of Feature Experts (MoFE) and Dynamicβ Mixup data augmentation to improve the performance of visual foundation models in OOD detection.
Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts
Ibtihel Amara (Google Research), Negar Rostamzadeh (Google Research)
GenerationData SynthesisDiffusion modelImageTextBenchmark
🎯 What it does: Evaluate and analyze the impact of text-to-image models on non-target concepts after concept erasure, and propose the EraseBench benchmark.
ERNet: Efficient Non-Rigid Registration Network for Point Sequences
Guangzhao He (State Key Laboratory of Computer Aided Design and Computer Graphics), Sida Peng (State Key Laboratory of Computer Aided Design and Computer Graphics)
OptimizationComputational EfficiencyTransformerPoint Cloud
🎯 What it does: This paper proposes an efficient sequence non-rigid registration network ERNet based on deformation maps, capable of real-time and accurate registration of sparse or partial point cloud sequences.
Error Recognition in Procedural Videos using Generalized Task Graph
Shih-Po Lee (Northeastern University), Ehsan Elhamifar (Northeastern University)
RecognitionSegmentationAnomaly DetectionLarge Language ModelVision Language ModelVideoText
🎯 What it does: A unified framework is proposed to jointly achieve temporal action segmentation and error recognition in process videos under the condition of error-free video training.
ESCNet:Edge-Semantic Collaborative Network for Camouflaged Object Detection
Sheng Ye (Xiamen University), Liujuan Cao (Xiamen University)
Object DetectionSegmentationTransformerImage
🎯 What it does: The ESCNet framework is proposed for camouflage object detection, utilizing dynamically coupled edge-texture perception for fine segmentation.
ESSENTIAL: Episodic and Semantic Memory Integration for Video Class-Incremental Learning
Jongseo Lee (Kyung Hee University), Jinwoo Choi (Kyung Hee University)
ClassificationRecognitionVideoBenchmark
🎯 What it does: A video category incremental learning (VCIL) method called ESSENTIAL is proposed, which alleviates catastrophic forgetting by combining temporally sparse experience memory with learnable semantic prompts.
Estimating 2D Camera Motion with Hybrid Motion Basis
Haipeng Li (University of Electronic Science and Technology of China), Shuaicheng Liu (University of Electronic Science and Technology of China)
TransformerOptical FlowImage
🎯 What it does: This paper proposes a new 2D camera motion estimation framework called CamFlow, which represents nonlinear camera motion by mixing physical foundations and random noise foundations, and uses a Laplace distribution-based mixed probability loss for training.
ETA: Efficiency through Thinking Ahead, A Dual Approach to Self-Driving with Large Models
Shadi Hamdan (Koc University), Fatma Guney (Koc University)
Autonomous DrivingComputational EfficiencyReinforcement LearningVideo
🎯 What it does: Proposes the ETA dual-system architecture, moving the computation of the large model to the previous frame and using asynchronous batch inference to achieve real-time autonomous driving control.
ETA: Energy-based Test-time Adaptation for Depth Completion
Younjoon Chung (Yale University), Alex Wong (Yale University)
Depth EstimationDomain AdaptationPoint Cloud
🎯 What it does: An energy-based test-time adaptive method (ETA) is proposed, which evaluates the distribution similarity of depth predictions by training an energy model, and updates only a lightweight adaptation layer during inference to minimize energy, achieving adaptation of the target domain depth completion model.
ETCH: Generalizing Body Fitting to Clothed Humans via Equivariant Tightness
Boqian Li (Westlake University), Yuliang Xiu (Westlake University)
SegmentationPose EstimationTransformerSupervised Fine-TuningPoint CloudMesh
🎯 What it does: This paper proposes a clothing-body fitting framework ETCH based on SE(3)-equivariant tightness vectors, which maps the clothing surface to the internal body, achieving precise fitting of the body in any pose and clothing.
ETVA: Evaluation of Text-to-Video Alignment via Fine-grained Question Generation and Answering
Kaisi Guan (Renmin University of China), Ruihua Song (Renmin University of China)
TransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: This paper proposes ETVA, an automated method for evaluating text-to-video (T2V) alignment through fine-grained question generation and answering, and constructs a dedicated evaluation benchmark called ETVABench.
Evading Data Provenance in Deep Neural Networks
Hongyu Zhu (Shanghai Jiao Tong University), Shi-Lin Wang (Shanghai Jiao Tong University)
ClassificationKnowledge DistillationAdversarial AttackTransformerLarge Language ModelVision Language ModelImageBiomedical Data
🎯 What it does: A unified framework for escaping data origin verification (Escaping DOV) is proposed, where task knowledge is transferred from a teacher model to a student model using unrelated OOD datasets after learning copyright data, thereby circumventing all DOV methods.
EvaGaussians: Event Stream Assisted Gaussian Splatting from Blurry Images
Wangbo Yu (Peking University), Yonghong Tian (Peking University)
RestorationData SynthesisGaussian SplattingImage
🎯 What it does: Combining event camera streams with traditional frame cameras, we train 3D Gaussian Splatting to recover high-quality 3D scenes from motion-blurred images and achieve novel view synthesis.
EVDM: Event-based Real-world Video Deblurring with Mamba
Zhijing Sun (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationOptical FlowImageVideo
🎯 What it does: This paper proposes an event-driven video deblurring framework EVDM based on Mamba, which can utilize long-term motion information of events to achieve high-quality video deblurring.
Event-aided Dense and Continuous Point Tracking: Everywhere and Anytime
Zhexiong Wan (Northwestern Polytechnical University), Gim Hee Lee (National University of Singapore)
Object TrackingTransformerOptical FlowImageVideo
🎯 What it does: A point tracking framework that combines event cameras with traditional frame images is proposed to achieve spatially dense continuous point trajectory tracking at any point in time.
Event-based Tiny Object Detection: A Benchmark Dataset and Baseline
Nuo Chen (National University of Defense Technology), Wei An (Nankai University)
Object DetectionConvolutional Neural NetworkPoint CloudBenchmark
🎯 What it does: A large-scale benchmark dataset for small target detection under event cameras, named EV-UAV, has been constructed, and a detection framework based on sparse event point clouds, called EV-SpSegNet, along with its spatiotemporal correlation loss, has been proposed.
Event-based Visual Vibrometry
Xinyu Zhou (Peking University), Boxin Shi (Peking University)
OptimizationConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowImageVideoMultimodalityAudio
🎯 What it does: A dual-camera system that combines event cameras and conventional cameras is used to achieve high frame rate capture and precise motion estimation of subtle vibrations, thereby reconstructing sound signals and material properties of objects.
Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction
Wenhao Xu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
RestorationSegmentationData SynthesisOptimizationGaussian SplattingMultimodalityBenchmark
🎯 What it does: A dynamic scene reconstruction framework is proposed that combines event cameras with deformable 3D Gaussian splatting, significantly enhancing reconstruction and rendering quality by utilizing the ultra-high temporal and frequency information of events.
Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene
Donggeun Lim (Seoul National University), Young Min Kim (Seoul National University)
GenerationRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringBenchmark
🎯 What it does: A framework based on large language models is proposed, capable of automatically generating natural interactive actions for multiple characters in 3D scenes, supporting dynamic story generation based on user free text instructions.
Event-guided HDR Reconstruction with Diffusion Priors
Yixin Yang (Peking University), Boxin Shi (Peking University)
RestorationGenerationDiffusion modelImage
🎯 What it does: This paper proposes an event-guided HDR reconstruction method based on diffusion models, utilizing the dynamic range information from event cameras and low dynamic range images to jointly generate high-quality color HDR images.
Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring
Taewoo Kim (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
Image TranslationRestorationConvolutional Neural NetworkVideoMultimodality
🎯 What it does: A unified framework is proposed to synchronously capture RGB frames and event information from low-light videos using event cameras, achieving low-light enhancement, motion deblurring, and high frame rate interpolation in an end-to-end manner.
EventUPS: Uncalibrated Photometric Stereo Using an Event Camera
Jinxiu Liang (Peking University), Boxin Shi (Peking University)
Image
🎯 What it does: This paper proposes a calibration-free photometric stereo method using event cameras, called EventUPS, to recover surface normal vectors and illumination trajectories from asynchronous event streams.
EVER: Exact Volumetric Ellipsoid Rendering for Real-time View Synthesis
Alexander Mai (University of California), Yinda Zhang (Google)
Neural Radiance FieldPoint Cloud
🎯 What it does: An Exact Volumetric Ellipsoid Rendering (EVER) method is proposed, which achieves real-time, non-bouncing view synthesis by precisely merging any number of overlapping constant-density ellipsoids through dual-intersection ray tracing.
Everything is a Video: Unifying Modalities through Next-Frame Prediction
G. Thomas Hudson (Durham University), Noura Al Moubayed (Durham University)
ClassificationObject DetectionTransformerImageVideoTextMultimodalityAudio
🎯 What it does: A framework is proposed to unify multimodal tasks into next-frame prediction, eliminating modality-specific encoders.
EVEv2: Improved Baselines for Encoder-Free Vision-Language Models
Haiwen Diao (Dalian University of Technology), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationData SynthesisTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: EVEv2.0 is proposed and implemented, a novel decoder-only vision-language model without a visual encoder, along with a complete training process for visual perception from scratch.
Evidential Knowledge Distillation
Liangyu Xiang (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
ClassificationKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes to treat classification probabilities as random variables, modeling the outputs of the teacher model using a second-order Dirichlet distribution, and achieving knowledge distillation for the student model through alignment of both the expectation (macro) and the distribution itself (micro) of this distribution.
EVOLVE: Event-Guided Deformable Feature Transfer and Dual-Memory Refinement for Low-Light Video Object Segmentation
Jong-Hyeon Baek (Chungnam National University), Yeong Jun Koh (Chungnam National University)
SegmentationTransformerVideoMultimodality
🎯 What it does: A multi-modal framework EVOLVE for low-light video object segmentation is proposed, which integrates event camera data with RGB images, utilizes event-driven deformable convolution for feature alignment, and iteratively optimizes segmentation results through a dual-memory Transformer.
EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment
Yufei Zhu (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
OptimizationRobotic IntelligenceDiffusion modelImage
🎯 What it does: Proposes an iteratively evolving grasp generation framework that continuously optimizes multi-finger robotic grasp postures based on user preferences.
EvRT-DETR: Latent Space Adaptation of Image Detectors for Event-based Vision
Dmitrii Torbunov (Brookhaven National Laboratory), Yonggang Cui (Brookhaven National Laboratory)
Object DetectionDomain AdaptationAutonomous DrivingRecurrent Neural NetworkTransformerImage
🎯 What it does: This paper proposes the I2EvDet framework, which transforms mainstream image detectors (such as RT-DETR) into object detection models capable of handling event camera data, ultimately resulting in EvRT-DETR, which achieves a record improvement in detection accuracy.
EVT: Efficient View Transformation for Multi-Modal 3D Object Detection
Yongjin Lee (ThorDrive Company), Sanghyun Kim (Seoul National University)
Object DetectionAutonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: Proposes the EVT framework, achieving multi-modal 3D object detection through adaptive sampling and adaptive projection (ASAP), as well as improved query initialization and geometry-aware cross-attention;
ExCap3D: Expressive 3D Scene Understanding via Object Captioning with Varying Detail
Chandan Yeshwanth (Technical University of Munich), Angela Dai (Technical University of Munich)
SegmentationGenerationTransformerVision Language ModelPoint Cloud
🎯 What it does: The ExCap3D model is proposed, which can generate multi-level textual descriptions at both the object and part levels for each object in 3D scans, ensuring consistency between the two levels of descriptions.
Explaining Human Preferences via Metrics for Structured 3D Reconstruction
Jack Langerman (Independent Researcher), Dmytro Mishkin (Hover Inc.)
Object DetectionSegmentationData SynthesisPoint Cloud
🎯 What it does: This study investigates wireframe comparison metrics in 3D structure reconstruction and validates the consistency of various metrics with human preferences through expert evaluations, proposing a set of unit tests and learning-based metrics.
Exploiting Diffusion Prior for Task-driven Image Restoration
Jaeha Kim (Seoul National University), Kyoung Mu Lee (Seoul National University)
ClassificationRestorationObject DetectionSegmentationDiffusion modelImage
🎯 What it does: This paper proposes a task-driven image restoration method based on diffusion models, called EDTR, aimed at enhancing the performance of high-level visual tasks such as classification, semantic segmentation, and object detection by restoring key details in low-quality images.
Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation
Seogkyu Jeon (Yonsei University), Hyeran Byun (Yonsei University)
SegmentationDomain AdaptationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A semantic segmentation model based on Domain-Aware Prompt Driven Mask Transformer (DPMFormer) is proposed for single-source domain generalization.
Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition
Meiqi Cao (Nanjing University of Science and Technology), Jinhui Tang (Nanjing Forestry University)
RecognitionSafty and PrivacyTransformerVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: Privacy-friendly multimodal action recognition on event camera data, utilizing an event reconstruction network to generate texture-rich pseudo-RGB frames and jointly encoding them with event stacked frames.
Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation
Tiankai Chen (Southwest Jiaotong University), Xun Xu (Institute for Infocomm Research)
Object DetectionAnomaly DetectionGraph Neural NetworkTransformerVision Language ModelPoint Cloud
🎯 What it does: A training-free graph score propagation framework GSP is proposed for OOV detection of 3D point clouds using a visual-language model (VLM);
ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors
Minsu Kim (Yonsei University), Seon Joo Kim (Yonsei University)
RestorationGenerationDiffusion modelGaussian SplattingImageVideo
🎯 What it does: A detectable scene reconstruction framework based on 3D Gaussian Splatting and video diffusion models—ExploreGS—has been constructed, capable of generating flawless high-quality images from any viewpoint.
Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing
Joonghyuk Shin (Seoul National University), Jaesik Park (Seoul National University)
Image TranslationGenerationTransformerPrompt EngineeringDiffusion modelImageMultimodality
🎯 What it does: A prompt-driven image editing method based on the MM-DiT attention mechanism is proposed, capable of achieving high-quality global and local edits.
Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation
I-Hsiang Chen (University of Washington), Sy-Yen Kuo (National Taiwan University)
SegmentationDomain AdaptationDiffusion modelImage
🎯 What it does: A PDAF framework is proposed, introducing a probabilistic diffusion alignment mechanism in semantic segmentation models to enhance domain generalization performance.
Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics
Taowen Wang (Rochester Institute of Technology), Ruixiang Tang (Rutgers University)
Adversarial AttackRobotic IntelligenceVision-Language-Action ModelMultimodality
🎯 What it does: This paper studies the adversarial vulnerability of visual-language-action (VLA) models in robotic execution and proposes patch-based undirected and directed attack methods that can significantly reduce task success rates in both digital and physical environments.
Exploring The Visual Feature Space for Multimodal Neural Decoding
Weihao Xia (University of Cambridge), Cengiz Oztireli (University of Cambridge)
ClassificationRecognitionGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodalityBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: Aiming at human brain fMRI signals, the VINDEX method is proposed, utilizing various visual feature spaces and a pre-trained multimodal large language model (MLLM) for zero-shot multi-level decoding of brain signals, achieving fine-grained textual descriptions, localization, and reasoning of visual stimuli.
Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement
Shuo Zhang (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: To address the issue of illumination consistency in low-light light field images, we propose the VCNet network, which includes a lighting estimation module and a recovery module with self-supervised view consistency loss, capable of adaptively enhancing low-light light field images in various scenes.