CVPR 2026 Papers — Page 3
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
APPO: Attention-guided Perception Policy Optimization for Video Reasoning
Henghui Du (Renmin University of China), Di Hu (Renmin University of China)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodality
🎯 What it does: This paper proposes an attention-guided perception policy optimization algorithm called APPO, aimed at enhancing the fine-grained perception capabilities of multimodal large language models in video reasoning tasks.
AR2-4FV: Anchored Referring and Re-identification for Long-Term Grounding in Fixed-View Videos
Teng Yan (Hong Kong University of Science and Technology), Bingzhuo Zhong (Hong Kong University of Science and Technology)
Object DetectionObject TrackingRetrievalTransformerVision Language ModelContrastive LearningVideoTextBenchmark
🎯 What it does: Propose a language-guided localization and re-identification framework AR-4FV for fixed-perspective long-duration videos, which can maintain identity consistency and generate bounding boxes for each frame even when the target disappears or reappears after long-term occlusion.
Ar2Can: An Architect and an Artist Leveraging a Canvas for Multi-Human Generation
Shubhankar Borse (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
GenerationData SynthesisTransformerLarge Language ModelReinforcement LearningDiffusion modelImageBenchmark
🎯 What it does: Propose a two-stage multi-human generation framework: first using Architect to predict layout, then using Artist for high-quality rendering with identity preservation.
ARC Is a Vision Problem!
Keya Hu (MIT), Kaiming He (MIT)
Image TranslationImageBenchmark
🎯 What it does: Treat ARC as an image-to-image translation problem, using canvas-based inputs and training on Vision Transformer or U-Net;
Arcadia: Toward a Full-Lifecycle Framework for Embodied Lifelong Learning
Minghe Gao (Zhejiang University), Yueting Zhuang (Bytedance Seed)
Robotic IntelligenceTransformerVision Language ModelGaussian SplattingSimultaneous Localization and MappingMultimodalityPoint CloudMesh
🎯 What it does: Construct the Arcadia lifelong learning closed-loop framework, integrating autonomous exploration data collection in real environments, generating editable simulation scenarios, sharing multimodal navigation and control representations, and achieving continuous improvement cycles through Sim-from-Real feedback mechanisms that transition from real to simulation and back to real.
Archon: A Unified Multimodal Model for Holistic Digital Human Generation
Chong Bao (Zhejiang University), Yinda Zhang (Google)
GenerationTransformerLarge Language ModelDiffusion modelImageVideoTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Developed Archon, a fully pre-trained unified multimodal model that supports cross-modal generation, understanding, and editing from any modality (description, script, speech, animation, semantic video, image, video).
ArchSym: Detecting 3D-Grounded Architectural Symmetries in the Wild
Hanyu Chen (Cornell University), Noah Snavely (Cornell University)
RecognitionTransformerImage
🎯 What it does: This paper constructs a large-scale architectural symmetry dataset named ArchSym and proposes a 3D symmetry detection model based on a single RGB image. The model utilizes predicted scene geometry through signed distance maps to achieve scale-consistent symmetry plane localization.
Are Image-to-Video Models Good Zero-Shot Image Editors?
Zechuan Zhang (Zhejiang University), Yi Yang (Zhejiang University)
GenerationPrompt EngineeringDiffusion modelImageVideoChain-of-Thought
🎯 What it does: This paper proposes IF-Edit, a zero-shot image editing framework that does not require fine-tuning, utilizing a pre-trained image-to-video diffusion model to accomplish instruction-driven editing tasks.
Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation
Yiwen Tang (Shanghai AI Laboratory), Bin Zhao (Shanghai AI Laboratory)
GenerationLarge Language ModelReinforcement LearningTextMeshBenchmark
🎯 What it does: This paper proposes and systematically evaluates the feasibility of applying reinforcement learning (RL) to text-to-3D autoregressive generation, and based on this, develops a novel hierarchical RL framework Hi-GRPO and an RL-enhanced model AR3D-R1; meanwhile, it introduces the MME-3DR benchmark focusing on implicit reasoning capabilities.
AREA3D: Active Reconstruction Agent with Unified Feed-Forward 3D Perception and Vision-Language Guidance
Tianling Xu (Southern University of Science and Technology), Hanspeter Pfister (Harvard University)
OptimizationGaussian SplattingPoint CloudMeshBenchmark
🎯 What it does: Proposes AREA3D, an active 3D reconstruction agent that integrates forward 3D perception and vision-language guidance to actively select the most informative next view under a limited perspective budget.
ARES: Unifying Asymmetric RGB-Event Stereo for Probabilistic Scene Flow Estimation
Jie Long Lee (National University of Singapore), Gim Hee Lee (National University of Singapore)
Object TrackingDepth EstimationAutonomous DrivingTransformerOptical FlowMultimodality
🎯 What it does: Propose a heterogeneous RGB-event stereo fusion framework named ARES, which jointly estimates optical flow, disparity, and scene flow.
ARGUS: Defending Against Multimodal Indirect Prompt Injection via Steering Instruction-Following Behavior
Weikai Lu (South China University Of Technology), Hao Peng (Beihang University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: This study proposes ARGUS, a defense framework against multi-modal indirect injection attacks (IPI) based on activation space control, which includes the construction of a cross-image, video, and audio IPI benchmark, controllability analysis of instruction-following behavior, activation steering, injection detection, and post-processing mechanisms.
ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning
Shengyuan Ding (Fudan University), Jiaqi Wang (Shanghai Innovation Institute)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed an agent-based multimodal reward model called ARM-Thinker, which can autonomously invoke tools such as image cropping and document retrieval during evaluation, and achieve verifiable judgments through a think-act-observe loop;
ARMFlow: AutoRegressive MeanFlow for Online 3D Human Reaction Generation
Zichen Geng (University of Western Australia), Ajmal Saeed Mian (University of Western Australia)
GenerationConvolutional Neural NetworkTransformerDiffusion modelFlow-based ModelAuto Encoder
🎯 What it does: This paper proposes an autoregressive framework called ARMFlow based on MeanFlow for online 3D human reaction generation, and designs an offline version named ReMFlow;
ART: Articulated Reconstruction Transformer
Zizhang Li (Reality Labs Research, Meta), Zhao Dong (Reality Labs Research, Meta)
RestorationTransformerNeural Radiance FieldImageMesh
🎯 What it does: Propose ART (Articulated Reconstruction Transformer), a Transformer-based feedforward model that can predict the complete 3D geometry, texture, and motion parameters of articulated objects using only sparse multi-state RGB images.
ArtHOI: Taming Foundation Models for Monocular 4D Reconstruction of Hand-Articulated-Object Interactions
Zikai Wang (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
Object TrackingSegmentationPose EstimationDepth EstimationOptimizationTransformerLarge Language ModelDiffusion modelVideoPoint Cloud
🎯 What it does: Achieve 4D reconstruction of hand-object interaction from monocular RGB videos by leveraging prior information from multiple base models and optimizing for unified alignment;
ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding
Shuo Cao (USTC), Yihao Liu (Shanghai AI Lab)
Large Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes ArtiMuse, a multimodal large language model capable of providing expert-level text commentary and precise aesthetic scores simultaneously.
Artiverse: A Diverse and Physically Grounded Dataset for Articulated Objects
Denys Iliash (Simon Fraser University), Angel X. Chang (Simon Fraser University)
Data SynthesisLarge Language ModelImageMeshBenchmark
🎯 What it does: Constructed a large-scale dataset named Artiverse consisting of approximately 5,402 highly physically realistic and interactive 3D man-made objects across 88 subcategories, achieved through a semi-automated annotation process enabling multi-dimensional labels including functional components, joint movements, material density, mass, etc.
ArtLLM: Generating Articulated Assets via 3D LLM
Penghao Wang (ShanghaiTech University), Jiayuan Gu (ShanghaiTech University)
GenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityPoint Cloud
🎯 What it does: Proposes ArtLLM, a framework that transforms 3D joint and part layout prediction into language modeling, enabling the rapid generation of high-quality, simulatable articulated 3D assets from single images or text.
ArtPro: Self-Supervised Articulated Object Reconstruction with Adaptive Integration of Mobility Proposals
Xuelu Li (Shandong University), Changhe Tu (Shandong University)
GenerationPose EstimationContrastive LearningGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a framework based on self-supervised 3D Gaussian Splatting that can reconstruct digital twins of multi-part articulated objects without relying on prior segmentation.
Asking like Socrates: Socrates helps VLMs understand remote sensing images
Run Shao (Central South University), Haifeng Li (Baidu Inc.)
RecognitionRetrievalTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelImageTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the RS-EoT iterative evidence retrieval reasoning framework to address the pseudo reasoning problem in remote sensing images.
AssemblyBench: Physics-Aware Assembly of Complex Industrial Objects
Danrui Li (Rutgers State University of New Jersey), Anoop Cherian (Mitsubishi Electric Research Laboratories)
Pose EstimationRobotic IntelligenceTransformerVision Language ModelImageTextMultimodalityPoint CloudBenchmarkPhysics Related
🎯 What it does: This paper proposes the AssemblyBench dataset and the AssemblyDyno model, which utilize multimodal manuals to predict the assembly sequence and 6-DoF trajectory of industrial components.
Assignment-Driven Hash Learning in a Hyper-Semantic Space for On-the-Fly Category Discovery
Kaibing Yang (Southeast University), Tingzhang Luo (City University of Hong Kong)
ClassificationRecognitionRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A two-stage framework is proposed for OCD tasks: first, a hyper-semantic space is constructed with prototype enhancement and OOV interpolation, and then flexible prototype allocation and binary hash regularization are performed in this space to address the cascade degradation from features to hashes and the problem of monopolization by known class spaces.
AstraNav-Memory: Contexts Compression for Long Memory
Junjun Hu (Amap, Alibaba Group), Mu Xu (Amap, Alibaba Group)
CompressionAutonomous DrivingConvolutional Neural NetworkTransformerReinforcement LearningVision Language ModelContrastive LearningImageVideoMultimodalityBenchmark
🎯 What it does: Proposed an image-centric long-term memory framework, AstraNav-Memory, which utilizes a visual context compression module to achieve 30 visual tokens per frame in the Qwen2.5-VL model, enabling the accommodation of hundreds of historical images within a single context and supporting lifelong navigation.
AsymLoc: Towards Asymmetric Feature Matching for Efficient Visual Localization
Mohammad Omama (University of Texas at Austin), Yelin Kim (Amazon)
Pose EstimationKnowledge DistillationImage
🎯 What it does: Proposed the AsymLoc framework, achieving efficient matching in visual localization tasks by using a large Teacher model to process database images offline, a lightweight Student model to process query images online, and aligning features for effective matching;
Asynchronous Temporal Modeling with Two-Agent Framework for Streaming Dense Video Captioning
Yolo Y. Tang (University Of Rochester), Chenliang Xu (University Of Rochester)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: Proposed an asynchronous dual-agent framework called Takusen, which uses a lightweight SMM (Oracle) to first detect event boundaries, followed by a large-scale video-language model (Listener) that generates precise captions upon receiving signals from the Oracle; introduced a 'Silence Token' to avoid threshold triggering, achieving truly real-time streaming video captioning;
AT-VLA: Adaptive Tactile Injection for Enhanced Feedback Reaction in Vision-Language-Action Models
Xiaoqi Li (Peking University), Hao Dong (Peking University)
Robotic IntelligenceVision-Language-Action ModelMultimodality
🎯 What it does: Introduce tactile perception into existing Vision-Language-Action (VLA) frameworks, and achieve closed-loop control for contact-rich tasks through adaptive tactile injection and a dual-stream tactile reaction mechanism.
AToken: A Unified Tokenizer for Vision
Jiasen Lu (Apple), Yinfei Yang (Apple)
ClassificationGenerationRetrievalTransformerImageVideoTextMultimodalityMesh
🎯 What it does: Proposed and implemented a unified visual tokenizer ATOKEN that can simultaneously achieve high-fidelity reconstruction and semantic understanding on images, videos, and 3D assets.
AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots
Likui Zhang (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
Robotic IntelligenceMixture of ExpertsVision-Language-Action ModelVideoTextBenchmark
🎯 What it does: In this paper, we propose AtomicVLA, a unified think-execute framework that leverages atomic skill abstraction and Skill-Guided Mixture-of-Experts (SG-MoE) to achieve long-horizon task planning and action execution, while supporting continual learning.
Attack for Defense: Adversarial Agents for Point Prompt Optimization Empowering Segment Anything Model
Xueyu Liu (Taiyuan University of Technology), Yongle Chen (Taiyuan University of Technology)
SegmentationAdversarial AttackTransformerReinforcement LearningPrompt EngineeringImageBiomedical Data
🎯 What it does: Proposed Point Prompt Defender (PPD), which automatically optimizes the point prompts of the Segment Anything Model through an attack-defense dual-agent reinforcement learning framework.
Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing
Baifeng Shi (UC Berkeley), Hongxu Yin (NVIDIA)
RecognitionComputational EfficiencyConvolutional Neural NetworkTransformerReinforcement LearningVision Language ModelVideo
🎯 What it does: Propose the AutoGaze module, which automatically selects multi-scale video patches to significantly reduce the computational cost of ViT and MLLM.
Attention Surgery: An Efficient Recipe to Linearize Your Video Diffusion Transformer
Mohsen Ghafoorian (Qualcomm AI Research), Amirhossein Habibian (Qualcomm AI Research)
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelVideo
🎯 What it does: This paper proposes the Attention Surgery framework, which can replace the self-attention in pre-trained video diffusion Transformers (VDM) with linear or hybrid attention, significantly improving computational efficiency while maintaining generation quality, without requiring retraining from scratch.
Attention-aware Inference Optimizations for Large Vision-Language Models with Memory-efficient Decoding
Fatih Ilhan (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
OptimizationComputational EfficiencyVision Language ModelImageVideo
🎯 What it does: Propose the AttentionPack framework, which improves memory efficiency and reduces latency during the inference phase of large-scale vision-language models by compressing the KV cache.
Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models
Katarzyna Zaleska (Warsaw University Of Technology), Kamil Deja (Warsaw University Of Technology)
GenerationExplainability and InterpretabilitySupervised Fine-TuningDiffusion modelContrastive LearningImageText
🎯 What it does: This paper proposes a hierarchical localization method based on linear probing to identify the layers in text-to-image diffusion models where implicit decisions (e.g., default fillings of attributes such as gender, age, and race) occur, and implements fine-grained intervention (ICM) for debiasing and controllable generation;
Attribute-Preserving Pseudo-Labeling for Diffusion-Based Face Swapping
Jiwon Kang (KAIST AI), Seungryong Kim (SAMSUNG)
Image TranslationGenerationDiffusion modelRectified FlowImage
🎯 What it does: This paper proposes the APPLE framework, a teacher-student framework based on diffusion models, which achieves a balance between identity transfer and attribute preservation in face swapping by utilizing attribute-preserving pseudo labels.
Attribution as Retrieval: Model-Agnostic AI-Generated Image Attribution
Hongsong Wang (Southeast University), Jie Gui (Southeast University)
RetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Designed and implemented a retrieval-based AI-generated image attribution framework called LIDA, which uses 'generated fingerprints' produced from low-bit planes as input. After unsupervised pre-training and few-shot adaptation, the framework can efficiently attribute and detect deepfakes for unknown generators.
Attribution-Guided Model Rectification of Unreliable Neural Network Behaviors
Peiyu Yang (University of Melbourne), Ajmal Mian (University of Western Australia)
Explainability and InterpretabilityAdversarial AttackImage
🎯 What it does: The study proposes a dynamic layer localization mechanism that utilizes rank-one model editing and attribution guidance to correct unreliable behaviors in models caused by backdoors, pseudo-correlations, or feature leakage.
Audio-sync Video Instance Editing with Granularity-Aware Mask Refiner
Haojie Zheng (Peking University), Xinlong Wang (Beijing Academy of Artificial Intelligence)
SegmentationGenerationTransformerLarge Language ModelAgentic AIVision-Language-Action ModelDiffusion modelVideoMultimodalityAudio
🎯 What it does: Proposes the AVI-Edit framework to achieve audio-synchronized video instance-level editing based on user-provided instance masks and text instructions;
AudioAvatar: Personalized Audio-driven Whole-body Talking Avatars
Seungeun Lee, Gyeong-Moon Park (Klleon Ai Research)
GenerationData SynthesisDiffusion modelContrastive LearningGaussian SplattingVideoMultimodalityAudio
🎯 What it does: This paper proposes an end-to-end single-image personalized full-body audio-driven conversational character model that can directly control the entire human expression and motion from audio.
AudioStory: Generating Long-Form Narrative Audio with Large Language Models
Yuxin Guo (University of Chinese Academy of Sciences), Wei Zou (University of Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelDiffusion modelFlow-based ModelAudio
🎯 What it does: This paper proposes AudioStory, an end-to-end framework that integrates large language models (LLM) with audio diffusion models (DiT) to generate long-form narrative audio, achieving instruction following, temporal coherence, and emotional consistency;
AURA: Multi-modal Shared Autonomy for Urban Navigation
Yukai Ma (University of California, Los Angeles), Bolei Zhou (University of California, Los Angeles)
Autonomous DrivingTransformerSupervised Fine-TuningVision-Language-Action ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose the AURA framework, decomposing urban roaming tasks into high-level human instructions (text, drawings, arrows) and low-level AI control, achieving multi-modal shared autonomy.
Authorize-on-Demand: Dynamic Authorization with Legality-Aware Intellectual Property Protection for VLMs
Lianyu Wang (Key Laboratory of Brain Machine Intelligence Technology Ministry of Education), Daoqiang Zhang (Key Laboratory of Brain Machine Intelligence Technology Ministry of Education)
ClassificationDomain AdaptationSafty and PrivacyVision Language ModelContrastive LearningImage
🎯 What it does: Propose a dynamic authorization VLM (Vision-Language Model) intellectual property protection framework named AoD-IP, which supports on-demand switching of authorization domains and enables legitimacy judgment.
AutoCut: End-to-end advertisement video editing based on multimodal discretization and controllable generation
Milton Zhou (Tsinghua University), Peng Jiang (Kuaishou Technology)
GenerationRetrievalConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Propose AutoCut, an end-to-end advertising video editing framework based on multimodal discretization and controllable generation;
AutoDebias: An Automated Framework for Detecting and Mitigating Backdoor Biases in Text-to-Image Models
Hongyi Cai (Universiti Malaya), Qingsong Wen (Squirrel Ai Learning)
GenerationSafty and PrivacyLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposed and implemented the AutoDebias framework, which can automatically detect and eliminate malicious backdoor biases in text-to-image models.
AutoRegressive Generation with B-rep Holistic Token Sequence Representation
Jiahao Li (Northwestern Polytechnical University), Yilei Shi (Northwestern Polytechnical University)
GenerationTransformerAuto EncoderMeshSequential
🎯 What it does: Proposed the BrepARG framework, which unifies the geometric and topological information of B-rep into an overall token sequence, and completes end-to-end generation through autoregressive Transformer.
AutoTraces: Autoregressive Trajectory Forecasting via Multimodal Large Language Models
Teng Wang (Southeast University), Ruize Wang (Southeast University)
Autonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalitySequentialBenchmarkChain-of-Thought
🎯 What it does: Propose AutoTraces, an autoregressive vision-language-trajectory prediction framework based on large language models, achieving socially compliant trajectory prediction through trajectory tokenization and automatic chain-of-thought (CoT) reasoning.
AV-Reasoner: Improving and Benchmarking Clue-Grounded Audio-Visual Counting for MLLMs
Lidong Lu (Nanjing University), Tong Lu (China Mobile Zijin Innovation Institute)
RecognitionObject DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: This work proposes a new benchmark for long-video multimodal counting called CG-AV-Counting, and designs a dual-mode evaluation (black-box + white-box) on this benchmark. Based on this, the authors propose the AV-Reasoner model, which utilizes GRPO reinforcement learning and hierarchical curriculum learning to progressively enhance the model's performance on tasks such as counting, localization, and reasoning.
AVA-Bench: Atomic Visual Ability Benchmark for Vision Foundation Models
Zheda Mai (Ohio State University), Wei-Lun Chao (Boston University)
Large Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: Proposes AVA-Bench, an evaluation benchmark that specifically decomposes and assesses the performance of visual foundation models on 14 atomic visual abilities (such as localization, counting, spatial reasoning, direction, depth, color, texture, emotion, OCR, action, fine-grained classification, object, and scene);
AVA-VLA: Improving Vision-Language-Action models with Active Visual Attention
Lei Xiao (LiAuto Inc), Xiaoyuan Yu (LiAuto Inc)
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelImageTextSequential
🎯 What it does: Proposes the AVA-VLA framework, which improves the decision-making performance of vision-language-action models in partially observable environments by introducing recursive states and active visual attention mechanisms.
Avatar Forcing: Real-Time Interactive Head Avatar Generation for Natural Conversation
Taekyung Ki (KAIST), Sung Ju Hwang (KAIST)
GenerationTransformerDiffusion modelAuto EncoderVideoMultimodalityAudio
🎯 What it does: Proposes Avatar Forcing, a real-time interactive head avatar generation framework capable of instantly generating natural and expressive avatar videos based on user audio, motion, and avatar audio;
AVATAR: Reinforcement Learning to See, Hear, and Reason Over Video
Yogesh Kulkarni (Arizona State University), Pooyan Fazli (Arizona State University)
OptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningVideoMultimodalityBenchmarkAudio
🎯 What it does: Aiming at multi-modal (audio-visual) long-sequence reasoning tasks, the AVATAR framework is proposed, improving the training method of GRPO and addressing issues such as low data efficiency, advantage disappearance, and uniform credit allocation.
AvatarPointillist: AutoRegressive 4D Gaussian Avatarization
Hongyu Liu (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerImagePoint Cloud
🎯 What it does: Proposes the AvatarPointillist framework based on autoregressive Transformer, which can generate animatable 4D Gaussian Avatar from a single portrait image in one go;
AVFakeBench: A Comprehensive Audio-Video Forgery Detection Benchmark for AV-LMMs
Shuhan Xia (Beijing University of Posts and Telecommunications), Zekun Li (Beijing University of Posts and Telecommunications)
ClassificationAnomaly DetectionTransformerLarge Language ModelMultimodalityBenchmark
🎯 What it does: Proposed AVFakeBench, the first benchmark for audio-visual forgery detection that comprehensively covers human and general scenarios, 11 real-world scenarios, 7 multimodal forgery types, and 4-level annotations;
AVGGT: Rethinking Global Attention for Accelerating VGGT
Xianbing Sun (Shanghai Jiao Tong University), Jianfu Zhang (Shanghai Jiao Tong University)
Computational EfficiencyTransformerImageVideo
🎯 What it does: Study and accelerate two multi-view 3D vision Transformer models, VGGT and π³, by proposing a two-step acceleration strategy without training.
AviaSafe: A Physics-Informed Data-Driven Model for Aviation Safety-Critical Cloud Forecasts
Zijian Zhu (Fudan University), Hao Li (Fudan University)
TransformerTime SeriesPhysics Related
🎯 What it does: Proposed and implemented the AviaSafe model, capable of predicting four cloud microphysical species (CIWC, CLWC, CRWC, CSWC) globally with a 6-hour interval and a delay of up to seven days, providing more detailed cloud phase forecasts for aviation safety.
AVION: Aerial Vision-Language Instruction from Offline Teacher to Prompt-Tuned Network
Yu Hu (University of British Columbia Okanagan), Mohsen Zardadi (TerraSense Analytics)
ClassificationRetrievalKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This study proposes a knowledge distillation framework called AVION for efficiently transferring large vision-language models to remote sensing image tasks;
AwareVLN: Reasoning with Self-awareness for Vision-Language Navigation
Wenxuan Guo (Tsinghua University), Jiwen Lu (Tsinghua University)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelMultimodalityChain-of-Thought
🎯 What it does: In the visual language navigation task, the AwareVLN framework is proposed, endowing the navigation agent with self-aware structured reasoning capabilities, enabling scene description, progress assessment, and next-step planning at critical nodes.
AXG-Reasoner: Error Detection and Explanation in Long Task Videos with Vision-Language Models
Shih-Po Lee (Northeastern University), Ehsan Elhamifar (Northeastern University)
Anomaly DetectionExplainability and InterpretabilityGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelVision-Language-Action ModelVideo
🎯 What it does: Propose the AXG-Reasoner framework, combining a frozen Vision-Language Model (VLM) with an automatically constructed Action Execution Graph (AXG) and a Time Action Segmentation (TAS) model, to detect and explain errors in long-task videos.
b-CLIP: Text-Conditioned Contrastive Learning for Multi-Granular Vision-Language Alignment
Fatimah Zohra (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose a multi-grained text conditional contrastive learning framework β-CLIP, which aligns hierarchically decomposed text (complete sentences, clauses, phrases) with visual image regions and introduces β-Contextualized Contrastive Alignment Loss (β-CAL) to regulate semantic overlap.
B$^3$-Seg: Camera-Free, Training-Free 3DGS Segmentation via Analytic EIG and Beta-Bernoulli Bayesian Updates
Hiromichi Kamata (Sony Group Corporation), Fuminori Homma (Sony Group Corporation)
SegmentationGaussian Splatting
🎯 What it does: Proposes B-Seg 3, an interactive 3D Gaussian Splatting (3DGS) segmentation method that does not require camera trajectories, training, or annotations, and supports open-vocabulary segmentation, achieving high-quality results within seconds;
BA-GS: Bayesian Adaptive Gaussian Splatting for SFM-Free 3D Reconstruction
Zhongjie Ma (Tianjin University), Changqing Zhang (Tianjin University)
GenerationOptimizationGaussian SplattingImage
🎯 What it does: Propose BA-GS, which leverages a Bayesian framework to handle uncertainty in SFM-free 3D Gaussian Splatting by employing VB-GMM for global initialization and adaptive Kalman filtering for local refinement, achieving more robust 3D reconstruction under sparse views.
BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models
Shengao Wang (Boston University), Boqing Gong (Boston University)
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the BabyVLM-V2 framework, which integrates infant audio-visual data to construct a cross-modal pre-training dataset spanning video, images, and dialogue. The framework pre-trains and instruction-fine-tunes based on LLaMA + ViT vision-language models, while designing the DevCV Toolbox evaluation suite aligned with NIH Baby Toolbox.
Back to Basics: Let Denoising Generative Models Denoise
Tianhong Li (MIT), Kaiming He (MIT)
GenerationTransformerDiffusion modelImage
🎯 What it does: The paper proposes a diffusion model that directly predicts clean images in pixel space (x-prediction) and implements it using a pure Vision Transformer (ViT) architecture (i.e., 'Just image Transformers', JiT). By training on large-sized pixel patches, the model can generate high-quality images at resolutions of ImageNet 256×256, 512×512, and even 1024×1024.
Back to Point: Exploring Point-Language Models for Zero-Shot 3D Anomaly Detection
Kaiqiang Li (Qilu University of Technology Shandong Academy of Sciences), Jin Wan (Qilu University of Technology Shandong Academy of Sciences)
Anomaly DetectionTransformerPrompt EngineeringContrastive LearningPoint Cloud
🎯 What it does: Propose the BTP framework, which utilizes pre-trained point language models to directly perform zero-shot 3D anomaly detection and localization on raw point clouds.
Back to Source: Open-Set Continual Test-Time Adaptation via Domain Compensation
Yingkai Yang (Shenzhen University), Hui Huang (Shenzhen University)
Domain AdaptationTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes an end-to-end visual prompt learning framework named DOCO for open continual test-time adaptation (OCTTA) scenarios, aiming to simultaneously address domain drift and unknown class impact.
Back to the Feature: Explaining Video Classifiers with Video Counterfactual Explanations
Chao Wang (King's College London), Luis C. Garcia-Peraza-Herrera (King's College London)
Explainability and InterpretabilityDiffusion modelAuto EncoderVideo
🎯 What it does: Proposes the BTTF framework based on diffusion models for generating adversarial causal explanations for videos (video CFEs).
BackSplit: The Importance of Sub-dividing the Background in Biomedical Lesion Segmentation
Rachit Saluja (Cornell University), Mert R. Sabuncu (Cornell University)
SegmentationConvolutional Neural NetworkBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: Propose the BackSplit method to improve small lesion segmentation by subdividing the background into auxiliary structures
BAgger: Backwards Aggregation for Mitigating Drift in Autoregressive Video Diffusion Models
Ryan Po (Stanford University), Gordon Wetzstein
GenerationTransformerDiffusion modelVideoTextBenchmark
🎯 What it does: Propose a self-supervised reverse aggregation (BAgger) method that corrects drift issues by using the generated trajectories from the reverse autoregressive video diffusion model.
Balanced Dataset Distillation via Modeling Multiple Visual Pattern Distribution
Guanghui Shi (Xidian University), Qixiang Wen (Xidian University)
Knowledge DistillationData-Centric LearningContrastive LearningImage
🎯 What it does: To address the pattern imbalance problem in dataset distillation, we propose the BPS (Balanced Pattern Selection) framework based on multi-pattern distribution modeling. It constructs a hierarchical semantic structure within classes, uniformly selecting samples from both central and edge patterns to form a balanced core set, and trains the network using knowledge distillation.
Balanced Hierarchical Contrastive Learning with Decoupled Queries for Fine-grained Object Detection in Remote Sensing Images
Jingzhou Chen (Nanjing University of Science and Technology), Liang Xiao (Nanjing University of Science and Technology)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: In fine-grained object detection of remote sensing images, a hierarchical label structure is utilized to propose balanced hierarchical contrastive learning and query decoupling strategies to enhance detection performance.
BALM: A Model-Agnostic Framework for Balanced Multimodal Learning under Imbalanced Missing Rates
Phuong-Anh Nguyen (VNU University of Engineering and Technology), Cam-Van Thi Nguyen (VNU University of Engineering and Technology)
ClassificationRecognitionMultimodality
🎯 What it does: Proposed a model-agnostic framework named BALM, combining a feature calibration module and a gradient rebalancing module to address the representation and optimization imbalance caused by missing rate imbalance (IMR) in multi-modal learning.
BAMI: Training-Free Bias Mitigation in GUI Grounding
Borui Zhang (Tsinghua University), Jiwen Lu (Tsinghua University)
Object DetectionLarge Language ModelPrompt EngineeringImage
🎯 What it does: Propose a training-agnostic reasoning method called BAMI to improve the accuracy of GUI localization models in high-resolution, element-dense interfaces;
BarbieGait: An Identity-Consistent Synthetic Human Dataset with Versatile Cloth-Changing for Gait Recognition
Qingyuan Cai (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
RecognitionGenerationData SynthesisVideoBenchmark
🎯 What it does: Proposed the BarbieGait synthetic dataset and the GaitCLIF model for handling clothing variations,
Basis-Oriented Low-rank Transfer for Few-Shot and Test-Time Adaptation
Junghwan Park (Telepix), Kookjin Lee (Arizona State University)
ClassificationDomain AdaptationImage
🎯 What it does: Construct an orthogonal spectral basis on top of pre-trained models, leveraging the singular directions from existing multi-task fine-tuned models, and achieve low-rank transfer by learning only diagonal coefficients in new tasks.
Batch Loss Score for Dynamic Data Pruning
Qing Zhou (Northwestern Polytechnical University), Qi Wang (Northwestern Polytechnical University)
OptimizationComputational EfficiencyData-Centric LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes Batch Loss Score (BLS), which utilizes the mean batch loss available at each step to infer sample importance through sample-specific exponential moving average (EMA), achieving lossless dynamic data pruning;
Batman: Benign Knowledge Alignment Through Malicious Null Space in Federated Backdoor Attack
Wenwen He (Wuhan University), Mang Ye (Nanyang Technological University)
Federated LearningAdversarial AttackImage
🎯 What it does: Propose a new backdoor attack method in federated learning called Batman, which can maintain a high attack success rate while improving stealthiness.
Bayesian Decomposition and Semantic Completion for Few-shot Semantic Segmentation
Guangchen Shi (Nanjing University), Tong Lu (Nanjing University)
SegmentationMeta LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a Bayesian probabilistic decomposition-based few-shot semantic segmentation framework, BPNet, which decomposes the task into three parts: prior, likelihood, and class consistency. It uses a lightweight SAM to generate structural prior, CALM to achieve joint binary classification inference for consistency and likelihood, and SCM to complete segmentation and fuse global semantics.
BD-Merging: Bias-Aware Dynamic Model Merging with Evidence-Guided Contrastive Learning
Yuhan Xie (Shanghai University of Finance and Economics), Chen Lyu (Shanghai University of Finance and Economics)
Domain AdaptationComputational EfficiencyTransformerContrastive LearningImage
🎯 What it does: Propose a model fusion framework named BD-Merging, which can dynamically adjust fusion weights under distribution shifts during testing, significantly enhancing the robustness and generalization ability of multi-task model fusion.
BDNet:Bio-Inspired Dual-Backbone Small Object Detection Network
Wenchao Guan (Guangxi University of Science and Technology), Xintao Pang (Macao Polytechnic University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Propose a bio-inspired dual-trunk network BDNet for small object detection in remote sensing images.
BEA-GS: BEyond RAdiance Supervision in 3DGS for Precise Object Extraction
Alessio Mazzucchelli (Arquimea Research Center), Adrian Penate-Sanchez (Universidad de las Palmas de Gran Canaria)
SegmentationGaussian SplattingImage
🎯 What it does: Improve the geometric structure of 3D Gaussian Splatting scenes to achieve more precise boundaries during object extraction.
BeautyGRPO: Aesthetic Alignment for Face Retouching via Dynamic Path Guidance and Fine-Grained Preference Modeling
Jiachen Yang (Sun Yat-sen University), Yanmei Fang (Sun Yat-sen University)
Image HarmonizationSupervised Fine-TuningReinforcement LearningVision Language ModelFlow-based ModelImage
🎯 What it does: Aligns the facial beauty enhancement process with human aesthetic preferences through reinforcement learning, proposing a Dynamic Path Guidance (DPG) mechanism that enhances detail fidelity and exploratory capabilities.
Benchmarking Endoscopic Surgical Image Restoration and Beyond
Jialun Pei (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
RestorationSegmentationDepth EstimationConvolutional Neural NetworkTransformerOptical FlowBiomedical DataBenchmark
🎯 What it does: Proposed and made public SurgClean—the first real-world multi-type endoscopic surgery image restoration benchmark dataset—and constructed a unified evaluation framework based on it. Systematic evaluation of 22 common restoration algorithms was conducted, further analyzing structural differences between surgical and natural images and their impacts on depth estimation and semantic segmentation.
Benchmarking PhD-Level Coding in 3D Geometric Computer Vision
Wenyi Li (Tsinghua University), Hao Zhao (Tsinghua University)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Construct and release GeoCodeBench, a PhD-level code generation benchmark for 3D geometry computation and vision algorithms. Tasks are sourced from the official code repositories of top conferences in 2025, and provide structured paper texts, masked functions, and high-quality unit tests to evaluate LLMs' ability to generate executable code.
Benchmarking Single-Factor Physical Video-to-Audio Generation
Tingle Li (UC Berkeley), Ming-Yu Liu (NVIDIA)
GenerationData SynthesisContrastive LearningVideoMultimodalityBenchmarkPhysics RelatedAudio
🎯 What it does: Proposes the FlatSounds benchmark, specifically designed to evaluate the causal response of video-to-audio generation models to physical factors, using contrastive video pairs and single-video mode experiments.
Best Segmentation Buddies for Image-Shape Correspondence
Itai Lang (University of Chicago), Rana Hanocka (University of Chicago)
SegmentationKnowledge DistillationTransformerImageMesh
🎯 What it does: Studied the segmentation-level correspondence problem between images and textureless 3D meshes, proposing the Best Segmentation Buddies (BSB) method to achieve cross-modal and cross-domain segmentation alignment.
Better than Average: Spatially-Aware Aggregation of Segmentation Uncertainty Improves Downstream Performance
Vanessa Emanuela Guarino (Max-Delbrueck-Center), Carsten T. Lüth
SegmentationAnomaly DetectionImageBiomedical DataAgriculture Related
🎯 What it does: This paper studies how to aggregate pixel-level uncertainty maps into image-level scores, and systematically evaluates the impact of different aggregation strategies on downstream tasks (OOD detection and failure detection).
Better, Stronger, Faster: Tackling the Trilemma in MLLM-based Segmentation with Simultaneous Textual Mask Prediction
Jiazhen Liu (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
SegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposed an All-mask prediction paradigm, where STAMP generates text during the dialogue phase and subsequently predicts the entire image's segmentation mask in one parallel step, achieving unification of semantic dialogue and pixel-level segmentation.
BEV-CAR: Enhancing Monocular Bird's Eye View Segmentation with Context-Aware Rasterization
Yixin Xiong (Chongqing University), Kai Liu (Chongqing University)
SegmentationDepth EstimationAutonomous DrivingTransformerImage
🎯 What it does: Propose a BEV-CAR framework based on depth perception and global feature fusion, and optimize BEV semantic segmentation during training using context-aware rasterization (ray-wise 1D rasterization);
BEV-SLD: Self-Supervised Scene Landmark Detection for Global Localization with LiDAR Bird's-Eye View Images
David Skuddis (University of Stuttgart), Norbert Haala (University of Stuttgart)
Pose EstimationAutonomous DrivingRepresentation LearningConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: Propose the BEV-SLD method, which learns scene-specific keypoints (scene landmarks) in LiDAR BEV images through self-supervised learning, achieving global localization without dense maps;
Beyond [CLS] Token: Query-Driven Token-Level Forgery Purification for Generalizable Deepfake Detection
Changshuo Wang (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
Anomaly DetectionTransformerContrastive LearningVideo
🎯 What it does: Proposes a query token-based deepfake detection framework called QTFP, which leverages learnable query tokens to aggregate local forgery evidence in the ViT backend and enhances detection performance through forgery probability contrastive learning and real attention alignment regularization.
Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning
Chun-Hsiao Yeh (FAIR at Meta), Fanyi Xiao (FAIR at Meta)
Representation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningVideoPoint Cloud
🎯 What it does: The GASP framework enhances the model's 3D spatial reasoning ability by injecting geometric priors into the Transformer layers of a vision-language model (VLM).
Beyond Appearance: Camouflaged Object Detection via Geometric Structure
Jinyu Han (Nanjing University of Science and Technology), Jinhui Tang (Nanjing Forestry University)
Object DetectionDepth EstimationConvolutional Neural NetworkTransformerMixture of ExpertsImageVideo
🎯 What it does: Propose a method to transform the monocular depth estimation model (Depth Anything v2) into a hidden object detection model called DepthSAM, achieving task alignment and geometric-semantic fusion through two core modules: the Sparse Mixture-of-Experts Adapter (SMEA) and the Geometric-Semantic Fusion Module (GSFM).
Beyond Binary Contrast: Modeling Continuous Skeleton Action Spaces with Transitional Anchors
Yingjie Feng (Harbin Institute of Technology), Zhuotao Tian (Harbin Institute of Technology)
RecognitionRetrievalRepresentation LearningContrastive LearningVideoGraph
🎯 What it does: Propose the TranCLR framework, combining action transitional anchors with multi-level geometric manifold calibration to improve continuity representation in self-supervised skeletal action recognition.
Beyond Caption-Based Queries in Video Moment Retrieval
David Pujol-Perich (University of Barcelona), Michael Wray (University of Bristol)
RetrievalTransformerLarge Language ModelVideoTextBenchmark
🎯 What it does: This paper proposes a Video Moment Retrieval method based on search queries and constructs three new search query benchmarks on existing datasets.
Beyond Duality: A Hybrid Framework of Leveraging Shared and Private Features for RGB-Event Object Detection
Keyao Wang (Hebei University Of Technology), Haiyong Chen (Hebei University Of Technology)
Object DetectionTransformerMultimodality
🎯 What it does: Propose the SPFD network, which utilizes frequency domain consistency to separate shared and private features, and adaptively fuses them in the TriAdapt Encoder and TriInject Decoder to achieve RGB-Event object detection;
Beyond Endpoints: Path-Centric Reasoning for Vectorized Off-Road Network Extraction
Wenfei Guan (Institute of Computing Technology Chinese Academy of Sciences), Yu Hu (Institute of Computing Technology Chinese Academy of Sciences)
SegmentationTransformerImage
🎯 What it does: Proposes an offline road network extraction method and a new dataset WildRoad, addressing topological errors in rugged terrain caused by traditional node-centric approaches.
Beyond Euclidean Gossip: KL-Barycentric Consensus on Heterogeneous and Imbalanced Images
Lu Xu (University of Hong Kong), Guosheng Yin (University of Hong Kong)
ClassificationSegmentationFederated LearningImageBiomedical Data
🎯 What it does: Propose a decentralized deep learning framework based on KL barycenter natural gradient variational inference, which can achieve model aggregation in non-i.i.d. and imbalanced sample environments.
Beyond Explicit Language: Plug-and-Play Visual-to-Linguistic Modeling Toward General Object Tracking
Kaiyang Lan (Zhejiang University of Technology), Dongyan Guo (Zhejiang University of Technology)
Object TrackingTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Designed a pluggable text inversion module that automatically generates pseudo-text descriptions from visual features and integrates implicit linguistic information into visual trackers through a multi-level semantic injection mechanism, achieving visual-language tracking without explicit language input.
Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models
Zhirong Shen (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Proposes a learnable linear prediction (LP²) framework for feature caching and prediction in Diffusion Transformers (DiT), enabling training-free inference acceleration.
Beyond Geometry: Artistic Disparity Synthesis for Immersive 2D-to-3D
Ping Chen (China Unicom), Shiguo Lian (China Unicom)
Image TranslationGenerationDepth EstimationConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes the Art3D framework, aiming to achieve conversion from single 2D images to 3D views through artistic difference synthesis, with a focus on learning global depth budget and local 'brushstroke' style in professional 3D movies;