ICCV 2025 Papers with AI Summaries
IEEE/CVF International Conference on Computer Vision · 2701 papers
→ ICCV 2025 papers with code (833)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 2701 ICCV 2025 papers by keyword, author or institution —
free trial on arXivSub.
"Principal Components" Enable A New Language of Images
Xin Wen (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes a 1D visual tokenizer SEMANTICIST based on PCA structure, achieving progressive image reconstruction from a single to hundreds of tokens;
2.5 Years in Class: A Multimodal Textbook for Vision-Language Pretraining
Wenqi Zhang (Zhejiang University), Lidong Bing (DAMO Academy Alibaba Group)
Data SynthesisRetrievalTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: A multimodal teaching video corpus based on instructional videos is proposed, and visual language models are pre-trained on it.
2D Gaussian Splatting-based Sparse-view Transparent Object Depth Reconstruction via Physics Simulation for Scene Update
Jeongyun Kim (Seoul National University), Ayoung Kim (Seoul National University)
SegmentationDepth EstimationGaussian SplattingImagePhysics Related
🎯 What it does: This paper proposes a transparent object sparse view depth reconstruction method (TRAN-D) based on 2D Gaussian Splatting, and updates the physical simulation of real-world scenes.
2HandedAfforder: Learning Precise Actionable Bimanual Affordances from Human Videos
Marvin Heidinger (Technische Universitat Darmstadt), Georgia Chalvatzaki (Technische Universitat Darmstadt)
Object DetectionSegmentationTransformerSupervised Fine-TuningVision Language ModelImageVideoBenchmark
🎯 What it does: By automatically extracting the interaction areas between hands and objects from human videos, a high-precision executable functional area for two-handed use is generated, and a 278K image dataset called 2HANDS is constructed.
3D Gaussian Map with Open-Set Semantic Grouping for Vision-Language Navigation
Jianzhe Gao (Zhejiang University), Wenguan Wang (Zhejiang University)
TransformerVision-Language-Action ModelGaussian SplattingPoint CloudBenchmark
🎯 What it does: This paper proposes a scene map (Egocentric Scene Map) constructed based on sparse differentiable 3D Gaussian atoms, and assigns open semantic labels to each atom through Open-Set Semantic Grouping, ultimately enhancing the decision-making ability of visual language navigation (VLN) agents using multi-level action prediction (Scene, View, Instance).
3D Gaussian Splatting Driven Multi-View Robust Physical Adversarial Camouflage Generation
Tianrui Lou (Sun Yat-Sen University), Xiaochun Cao (Sun Yat-Sen University)
GenerationAutonomous DrivingAdversarial AttackGaussian SplattingImage
🎯 What it does: A physical adversarial attack framework based on 3D Gaussian Splatting (PGA) has been developed, capable of generating multi-view robust adversarial camouflage for arbitrary targets in both digital and real environments.
3D Mesh Editing using Masked LRMs
Will Gao (University of Chicago), Nikolaos Sarafianos (Meta Reality Labs)
GenerationOptimizationTransformerMesh
🎯 What it does: A three-dimensional mesh editing method based on a conditional large reconstruction model (Masked LRM) is proposed, which randomly occludes three-dimensional regions in multi-view rendering from the input perspective and uses a single edited image as a condition, allowing the model to complete the editing of three-dimensional shapes in a single forward inference.
3D Test-time Adaptation via Graph Spectral Driven Point Shift
Xin Wei (Xidian University), Nannan Wang (Xidian University)
ClassificationDomain AdaptationGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a 3D point cloud test-time adaptation method based on spectral frequency domain (GSDTTA), which achieves online adaptation of target domain point clouds by learning adjustments to low-frequency coefficients in the spectral frequency domain. It also introduces a feature spectral self-training strategy that updates both the input point cloud and model parameters at each iteration.
3D-MOOD: Lifting 2D to 3D for Monocular Open-Set Object Detection
Yung-Hsu Yang (ETH Zurich), Zuria Bauer (ETH Zurich)
Object DetectionTransformerPoint Cloud
🎯 What it does: An end-to-end 3D-MOOD model is proposed, which elevates 2D detection results to 3D through a 3D bounding box head, supporting monocular open-set object detection.
3DGraphLLM: Combining Semantic Graphs and Large Language Models for 3D Scene Understanding
Tatiana Zemskova (AIRI), Dmitry Yudin (AIRI)
Object DetectionSegmentationComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPoint CloudGraph
🎯 What it does: Proposes 3DGraphLLM, which embeds the semantic relationships of 3D scene graphs into LLM to accomplish three-dimensional audiovisual language tasks.
3DGS-LM: Faster Gaussian-Splatting Optimization with Levenberg-Marquardt
Lukas Höllein (Technical University of Munich), Matthias Nießner (Technical University of Munich)
OptimizationComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: Improved the optimization process of 3D Gaussian Splatting (3DGS) by replacing the original Adam optimizer with a custom Levenberg-Marquardt iterator to accelerate scene reconstruction.
3DRealCar: An In-the-wild RGB-D Car Dataset with 360-degree Views
Xiaobiao Du (University of Technology Sydney), Xin Yu (Li Auto Inc.)
Object DetectionSegmentationGenerationDiffusion modelImagePoint CloudBenchmark
🎯 What it does: This paper constructs and publicly releases the first large-scale 3D vehicle dataset in a real environment, 3DRealCar, which includes 2,500 vehicles, approximately 200 360-degree RGB-D views, real-size point clouds, and 13 categories of body segmentation annotations, collected under three lighting conditions. Subsequently, benchmark evaluations are conducted on multi-tasks such as 2D detection, body segmentation, neural field reconstruction, diffusion-based view synthesis, and single-image 3D generation based on this dataset.
3DSRBench: A Comprehensive 3D Spatial Reasoning Benchmark
Wufei Ma (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
RecognitionData SynthesisTransformerLarge Language ModelImageTextBenchmark
🎯 What it does: This paper presents 3DSRBench, a dataset for 3D spatial reasoning visual question answering that contains 2,772 manually annotated entries, covering four main problem types: height, position, direction, and multi-object relationships.
4D Gaussian Splatting SLAM
Yanyan Li (Hangzhou Dianzi University), Federico Tombari (Google)
Object TrackingPose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingOptical FlowPoint Cloud
🎯 What it does: A 4D Gaussian Splatting SLAM system is proposed, capable of simultaneously estimating camera trajectories and constructing dynamic Gaussian radiance fields in dynamic scenes.
4D Visual Pre-training for Robot Learning
Chengkai Hou (Peking University), Huazhe Xu (Tsinghua University)
Robotic IntelligenceReinforcement LearningDiffusion modelPoint Cloud
🎯 What it does: A 4D visual pre-training framework called FVP is proposed, which utilizes the previous frame's point cloud and robot action information to predict the next frame's point cloud through a conditional diffusion model, thereby learning general 3D visual representations that are subsequently used to enhance tasks such as robot grasping and manipulation.
4D-Bench: Benchmarking Multi-modal Large Language Models for 4D Object Understanding
Wenxuan Zhu (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
RecognitionObject DetectionData SynthesisTransformerLarge Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes the 4D-Bench benchmark, aimed at systematically evaluating the understanding capabilities of multimodal large language models (MLLMs) regarding 4D (3D objects evolving over time), including two main tasks: question answering (QA) and captioning with multi-view spatial-temporal reasoning.
4DSegStreamer: Streaming 4D Panoptic Segmentation via Dual Threads
Ling Liu (Tsinghua University), Li Yi (Tsinghua University)
Object DetectionSegmentationAutonomous DrivingRecurrent Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes 4DSegStreamer, a dual-thread system that implements 4D point cloud streaming panoramic segmentation, balancing real-time performance and accuracy.
6DOPE-GS: Online 6D Object Pose Estimation using Gaussian Splatting
Yufeng Jin (Technische Universitat Darmstadt), Georgia Chalvatzaki (Technische Universitat Darmstadt)
Pose EstimationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes an online 6D pose estimation and reconstruction framework called 6DOPE-GS based on 2D Gaussian distribution projection rendering.
7DGS: Unified Spatial-Temporal-Angular Gaussian Splatting
Zhongpai Gao (United Imaging Intelligence), Ziyan Wu (United Imaging Intelligence)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: A unified seven-dimensional Gaussian scattering framework (7DGS) is proposed, capable of real-time rendering of dynamic scenes that are dependent on space, time, and viewpoint.
A Conditional Probability Framework for Compositional Zero-shot Learning
Peng Wu (Shandong University), Wenguan Wang (Zhejiang University)
RecognitionTransformerContrastive LearningImageText
🎯 What it does: This paper proposes a Conditional Probability Framework (CPF) that decomposes joint probabilities into object likelihood and conditional attribute likelihood, and utilizes text-enhanced object features to guide attribute learning, thereby better capturing the semantic constraints and contextual dependencies between attributes and objects, achieving stronger compositional zero-shot recognition.
A Constrained Optimization Approach for Gaussian Splatting from Coarsely-posed Images and Noisy Lidar Point Clouds
Jizong Peng (dConstruct Robotics), Angela Yao (National University of Singapore)
Autonomous DrivingOptimizationGaussian SplattingSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: A constrained optimization framework is proposed, which utilizes multi-camera SLAM + LiDAR sensor data to jointly optimize camera intrinsic parameters, extrinsic parameters, and a 3D Gaussian splatting scene in the absence of precise SfM camera poses and dense point clouds, achieving high-quality static scene reconstruction.
A Differentiable Wave Optics Model for End-to-End Computational Imaging System Optimization
Chi-Jui Ho (University of California, San Diego), Nicholas Antipa (University of California, San Diego)
RestorationOptimizationImage
🎯 What it does: A differentiable wave optics simulator is proposed, combining ray tracing and Rayleigh-Sommerfeld integrals to achieve accurate modeling of diffraction and aberrations in complex optical systems, and is used for end-to-end optical and algorithmic co-optimization.
A Framework for Double-Blind Federated Adaptation of Foundation Models
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Michigan State University)
Federated LearningSafty and PrivacyKnowledge DistillationTransformerImage
🎯 What it does: This paper proposes the BlindFed framework, which enables multiple data owners to collaboratively adapt large foundational models in a double-blind federated manner without sharing models or data with the learning service provider.
A Good Teacher Adapts Their Knowledge for Distillation
Chengyao Qian (Monash University), Mehrtash Harandi (Monash University)
Knowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: In this paper, the author first systematically analyzes the capacity gap problem in knowledge distillation, revealing that the significant mismatch in intra-class distribution between the teacher and student is the fundamental reason for the student's performance being lower than that of a medium teacher. Subsequently, an Adapted Intra-class Distribution (AID) method is proposed, which fine-tunes the teacher's intra-class distribution, allowing the teacher's output distribution to better align with the student's capabilities without retraining the teacher, thereby significantly improving the distillation effect.
A Hidden Stumbling Block in Generalized Category Discovery: Distracted Attention
Qiyu Xu (Yunnan Normal University), Yonghang Tai (Yunnan Normal University)
ClassificationRecognitionTransformerImage
🎯 What it does: Proposes the Attention Focusing (AF) module, which utilizes two components, TIME and TAP, to adaptively prune irrelevant tokens in the Vision Transformer, thereby reducing distracting attention and enhancing feature quality in Generalized Category Discovery (GCD).
A Hyperdimensional One Place Signature to Represent Them All: Stackable Descriptors For Visual Place Recognition
Connor Malone (Queensland University of Technology), Michael Milford (Queensland University of Technology)
RecognitionAutonomous DrivingSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes a Hyperdimensional One Place Signatures (HOPS) framework based on high-dimensional computation, which fuses VPR descriptors of the same location under different conditions to enhance the robustness of visual localization.
A Lesson in Splats: Teacher-Guided Diffusion for 3D Gaussian Splats Generation with 2D Supervision
Chensheng Peng, Or Litany
RestorationGenerationData SynthesisDiffusion modelGaussian SplattingImageVideoPoint CloudStochastic Differential Equation
🎯 What it does: A framework is proposed that utilizes 2D supervised training for 3D diffusion models, using a pre-trained deterministic 3D reconstruction model as a 'noise teacher' and supervising the multi-step denoising process through rendering loss.
A Linear N-Point Solver for Structure and Motion from Asynchronous Tracks
Hang Su (ShanghaiTech University), Laurent Kneip (ShanghaiTech University)
Pose EstimationOptimizationSimultaneous Localization and MappingTime Series
🎯 What it does: A linear N-point solver is designed to simultaneously recover 3D structure and camera linear motion using asynchronous timestamped point observations.
A Plug-and-Play Physical Motion Restoration Approach for In-the-Wild High-Difficulty Motions
Youliang Zhang (Tsinghua University), Xiu Li (Tsinghua University)
RestorationPose EstimationTransformerReinforcement LearningDiffusion modelVideo
🎯 What it does: A pluggable physical motion recovery framework is proposed, which includes a mask condition correction module (MCM) for correcting motion defects in video capture and a physics-based motion transfer module (PTM).
A Quality-Guided Mixture of Score-Fusion Experts Framework for Human Recognition
Jie Zhu (Michigan State University), Xiaoming Liu (Michigan State University)
RecognitionMixture of ExpertsMultimodality
🎯 What it does: A QME (Quality-guided Mixture of Score-fusion Experts) framework is proposed, which achieves score-level fusion through quality-aware expert mixtures in full-body biometrics, and directly optimizes verification and open-set metrics using score triplet loss.
A Real-world Display Inverse Rendering Dataset
Seokjun Choi (POSTECH), Seung-Hwan Baek (Meta)
ImageBenchmark
🎯 What it does: A real-world dataset for display inverse rendering is proposed, and a display-camera imaging system is constructed and calibrated to capture objects with various geometric and reflective properties.
A Recipe for Generating 3D Worlds from a Single Image
Katja Schwarz (Meta Reality Labs), Peter Kontschieder
GenerationData SynthesisDepth EstimationDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a method for rapidly generating immersive 3D worlds from a single image. It first obtains a 360° panoramic image through zero-shot 2D panoramic synthesis, then elevates it to 3D space using depth estimation, and finally constructs a real-time renderable 3D scene through point cloud conditional image inpainting and Gaussian Splatting.
A Simple yet Mighty Hartley Diffusion Versatilist for Generalizable Dense Vision Tasks
Qi Bi (Westlake University), Yefeng Zheng (Yale University)
RestorationSegmentationDepth EstimationDomain AdaptationTransformerDiffusion modelImage
🎯 What it does: A diffusion model called HarDiff based on Discrete Hartley Transform (DHT) is proposed for general dense visual tasks (semantic segmentation, depth estimation, dehazing, etc.), which utilizes low-frequency information for training and high-frequency information for sampling, thereby achieving cross-domain generalization.
A Structure-aware and Motion-adaptive Framework for 3D Human Pose Estimation with Mamba
Ye Lu (Nanyang Technological University), Kim-Hui Yap (Nanyang Technological University)
Pose EstimationVideoBenchmark
🎯 What it does: This paper proposes a structure-aware and motion-adaptive framework SAMA based on Mamba, which includes a structure-aware state integrator and a motion-adaptive state modulator, aimed at preserving joint topology information while learning independent motion scales for different joints in 3D pose enhancement tasks.
A Tiny Change, A Giant Leap: Long-Tailed Class-Incremental Learning via Geometric Prototype Alignment
Xinyi Lai (Fuzhou University), Yuanlong Yu (Fuzhou University)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: A geometric prototype alignment (GPA) method is proposed to address the issues of gradient competition and catastrophic forgetting in long-tail category incremental learning.
A Token-level Text Image Foundation Model for Document Understanding
Tongkun Guan (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RecognitionSegmentationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper develops the TokenIT dataset, the TokenFD text-image foundational model, and the TokenVL multimodal large language model for fine semantic understanding and reasoning of text images such as documents, tables, and charts.
A Unified Framework for Industrial Cel-Animation Colorization with Temporal-Structural Awareness
Xiaoyi Feng (Link-To), Kaifeng Zou (Link-To)
Image TranslationTransformerOptical FlowImageVideo
🎯 What it does: A unified colorization framework for industrial cel animation is proposed, achieving paragraph-level precise coloring of line drawings through multi-keyframe temporal information and structural matching.
A Unified Framework for Motion Reasoning and Generation in Human Interaction
Jeongeun Park (Korea University), Sangdoo Yun (Naver AI Lab)
GenerationRetrievalTransformerLarge Language ModelDiffusion modelTextMultimodality
🎯 What it does: Developed the MoLaM interactive motion-language unified model, capable of understanding, generating, editing, and reasoning about two-person interactive movements in multi-turn dialogues.
A Unified Framework to BRIDGE Complete and Incomplete Deep Multi-View Clustering under Non-IID Missing Patterns
Xiaorui Jiang (University of Science and Technology of China), Yong Liao (University of Science and Technology of China)
OptimizationAuto EncoderContrastive LearningMultimodality
🎯 What it does: A unified BRIDGE framework is proposed to bridge complete multi-view clustering (DMVC) and incomplete multi-view clustering (DIMVC), taking into account different multi-view interaction strategies and addressing the issue of non-IID missing patterns.
A Unified Interpretation of Training-Time Out-of-Distribution Detection
Xu Cheng (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: Provide a unified explanation of OOD detection methods during training, verify the dominant role of high-order interactions in OOD detection, and explain the reasons for the difficulty in detecting near OOD samples from the perspective of interactions.
A View-consistent Sampling Method for Regularized Training of Neural Radiance Fields
Aoxiang Fan (École Polytechnique Fédérale de Lausanne), Pascal Fua (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisOptimizationKnowledge DistillationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: A sampling method based on view-consistent distribution (View-consistent Sampling) and depth-pushing loss (Depth-pushing Loss) is proposed for implicit regularization of NeRF training, enhancing scene reconstruction and new view rendering quality under sparse viewpoints.
A Visual Leap in CLIP Compositionality Reasoning through Generation of Counterfactual Sets
Zexi Jia (WeChat AI Tencent Inc), Jie Zhou (WeChat AI Tencent Inc)
GenerationData SynthesisRetrievalTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper automatically generates adversarial image-text pairs through a block diffusion method, enhancing the combinatorial reasoning ability of VLM.
A0: An Affordance-Aware Hierarchical Model for General Robotic Manipulation
Rongtao Xu (Southern University of Science and Technology), Xiaodan Liang (Southern University of Science and Technology)
Robotic IntelligenceTransformerDiffusion modelMultimodalityOrdinary Differential Equation
🎯 What it does: This paper presents A0, a hierarchical, object-oriented affinity perception diffusion model designed to address the issues of spatial affinity understanding and action execution in robotic operations.
A3GS: Arbitrary Artistic Style into Arbitrary 3D Gaussian Splatting
Zhiyuan Fang (Zhejiang University), Yuchi Huo (Zhejiang Lab)
Image TranslationGenerationGraph Neural NetworkAuto EncoderGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes A3GS, a feed-forward network based on GCN autoencoders and AdaIN, achieving zero-shot instant style transfer of any image style in any 3D Gaussian Splatting scene.
AAA-Gaussians: Anti-Aliased and Artifact-Free 3D Gaussian Rendering
Michael Steiner, Markus Steinberger
GenerationComputational EfficiencyGaussian SplattingImageBenchmark
🎯 What it does: Improved 3D Gaussian projection rendering, proposing the use of complete 3D Gaussian evaluation throughout the entire rendering process to eliminate visual defects such as aliasing, projection distortion, jitter, and popping.
ACAM-KD: Adaptive and Cooperative Attention Masking for Knowledge Distillation
Qizhen Lan (University of Alabama at Birmingham), Qing Tian (University of Alabama at Birmingham)
Object DetectionSegmentationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes Adaptive student-teacher Cooperative Attention Masking for Knowledge Distillation (ACAM-KD), used for model distillation in dense visual prediction tasks.
Accelerate 3D Object Detection Models via Zero-Shot Attention Key Pruning
Lizhen Xu (Xi'an Jiaotong University), Shanmin Pang (Xi'an Jiaotong University)
Object DetectionAutonomous DrivingComputational EfficiencyTransformerReinforcement LearningPoint Cloud
🎯 What it does: A zero-shot, no-training runtime pruning method tgGBC is proposed to dynamically prune low-importance keys in the Transformer decoder of 3D object detection models, thereby accelerating inference.
Accelerating Diffusion Sampling via Exploiting Local Transition Coherence
Shangwen Zhu (Shanghai Jiao Tong University), Fan Cheng (Shanghai Jiao Tong University)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageVideoText
🎯 What it does: A training-independent diffusion sampling acceleration method called LTC-Accel is proposed, which utilizes the local transition consistency of adjacent sampling steps to approximate the transition operator of the current step, thereby reducing computational load.
Accelerating Diffusion Transformer via Gradient-Optimized Cache
Junxiang Qiu (University of Science and Technology of China), Yanbin Hao (Hefei University of Technology)
GenerationData SynthesisOptimizationComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: Proposes the Gradient-Optimized Cache (GOC) scheme, which uses gradient caching and gradient optimization to reduce cache errors in Diffusion Transformer (DiT) sampling, significantly improving generation speed and quality;
AccidentalGS: 3D Gaussian Splatting from Accidental Camera Motion
Mao Mao (Zhejiang University), Zhaopeng Cui (Zhejiang University)
Data SynthesisPose EstimationGaussian SplattingOptical FlowImage
🎯 What it does: Proposes the AccidentalGS method, achieving pose-free neural 3D modeling and novel viewpoint synthesis from images with accidental camera motion.
ACE-G: Improving Generalization of Scene Coordinate Regression Through Query Pre-Training
Leonard Bruns (KTH Royal Institute of Technology), Eric Brachmann (Niantic Spatial)
Object DetectionPose EstimationTransformerContrastive LearningImage
🎯 What it does: Improved the positioning performance of the scene coordinate regression model under different perspectives and lighting conditions.
Achieving More with Less: Additive Prompt Tuning for Rehearsal-Free Class-Incremental Learning
Haoran Chen (Fudan University), Yu-Gang Jiang (Fudan University)
ClassificationRecognitionComputational EfficiencyTransformerPrompt EngineeringImage
🎯 What it does: A lightweight incremental learning method called Additive Prompt Tuning (APT) is proposed, which achieves multi-task learning in ViT by adding a small number of learnable prompts to the key-value vectors of the CLS token, thereby completing class incremental learning in a scenario without storage replay.
Acknowledging Focus Ambiguity in Visual Questions
Chongyan Chen (University of Texas at Austin), Danna Gurari (University of Colorado Boulder)
ClassificationRecognitionSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: The VQ-FocusAmbiguity dataset has been constructed and made publicly available, specifically annotating focal ambiguity in visual questions and evaluating model performance on ambiguity recognition and focal localization tasks.
Activation Subspaces for Out-of-Distribution Detection
Barış Zöngür (TU Darmstadt), Stefan Roth (hessian.AI)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a post-training OOD detection method named ActSub, which utilizes singular value decomposition of the classification head weight matrix to decompose network activations into deterministic and non-deterministic subspaces, and calculates similarity and energy scores separately in both, ultimately merging them to obtain an OOB discrimination score.
Active Learning Meets Foundation Models: Fast Remote Sensing Data Annotation for Object Detection
Marvin Burges (TU Wien), Dalton Lunga (Oak Ridge National Laboratory)
Object DetectionImage
🎯 What it does: A real-time active learning framework is proposed, utilizing the base model SAM to generate masks combined with detection boxes, achieving efficient labeling of semi-automated remote sensing image target detection datasets.
Active Membership Inference Test (aMINT): Enhancing Model Auditability with Multi-Task Learning.
Daniel DeAlcala (Universidad Autonoma de Madrid), Javier Ortega-Garcia (Universidad Autonoma de Madrid)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies an active membership inference testing method called aMINT, which simultaneously trains the original auditing model and the MINT model through multi-task learning to detect whether the data has been used for training.
Active Perception Meets Rule-Guided RL: A Two-Phase Approach for Precise Object Navigation in Complex Environments
Liang Qin (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Object DetectionRobotic IntelligenceReinforcement LearningImage
🎯 What it does: This paper proposes a two-stage object navigation framework called APRR, which first uses rule-guided reinforcement learning to efficiently explore unknown environments, and then employs active perception reinforcement learning to refine the precise docking position at the target.
AcZeroTS: Active Learning for Zero-shot Tissue Segmentation in Pathology Images
Jiao Tang (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)
SegmentationTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: A zero-shot histopathological image segmentation framework based on active learning, AcZeroTS, is proposed, which can enhance the segmentation performance for both seen and unseen tissue types with only a small amount of labeled data for seen categories.
AD-GS: Object-Aware B-Spline Gaussian Splatting for Self-Supervised Autonomous Driving
Jiawei Xu (Nankai University), Jian Yang (Nankai University)
Autonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: A completely self-supervised Gaussian splatting-based autonomous driving scene rendering framework AD-GS has been constructed, capable of achieving high-quality free-viewpoint rendering without 3D annotations.
AdaDCP: Learning an Adapter with Discrete Cosine Prior for Clear-to-Adverse Domain Generalization
Qi Bi (Wuhan University), Gui-Song Xia (Wuhan University)
SegmentationDomain AdaptationAutonomous DrivingTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes AdaDCP, an adapter that combines discrete cosine transform (DCT) priors to achieve domain generalization for semantic segmentation of Vision Foundation Models (VFM) in various adverse weather scenarios, using only clear scene data for fine-tuning.
AdaDrive: Self-Adaptive Slow-Fast System for Language-Grounded Autonomous Driving
Ruifei Zhang (Chinese University of Hong Kong), Guanbin Li (Sun Yat-sen University)
Autonomous DrivingTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: AdaDrive achieves on-demand activation and dynamic fusion of LLM in autonomous driving through an adaptive slow-fast architecture, enhancing decision quality.
AdaHuman: Animatable Detailed 3D Human Generation with Compositional Multiview Diffusion
Yangyi Huang (NVIDIA), Umar Iqbal (NVIDIA)
GenerationData SynthesisPose EstimationDiffusion modelGaussian SplattingImageVideoStochastic Differential Equation
🎯 What it does: This study presents AdaHuman, which can generate high-quality, animatable 3D Gaussian Splat (3DGS) avatars from a single real-world image and supports arbitrary pose reshaping.
Adapt Foundational Segmentation Models with Heterogeneous Searching Space
Li Yi (Shanghai Jiao Tong University), Guannan Jiang (Contemporary Amperex Technology Co., Limited)
SegmentationDomain AdaptationReinforcement LearningGenerative Adversarial NetworkImage
🎯 What it does: An 'augmenting-to-adapt' framework is proposed, which utilizes a small number of labeled samples to search for optimal image enhancement strategies across multiple domains, achieving domain adaptation with zero-shot pre-trained segmentation models (such as SAM).
Adapting In-Domain Few-Shot Segmentation to New Domains without Source Domain Retraining
Qi Fan (Nanjing University), Yang Gao (Mohamed bin Zayed University of Artificial Intelligence)
SegmentationDomain AdaptationImageMagnetic Resonance Imaging
🎯 What it does: To address the cross-domain few-shot segmentation problem, we propose an adaptive structural adjustment of the pre-trained FSS model during inference using a small number of labeled support samples, enabling cross-domain transfer without retraining on the source domain.
Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision
Xiao Fang (Carnegie Mellon University), Fernando De La Torre (Carnegie Mellon University)
Object DetectionDomain AdaptationDiffusion modelImage
🎯 What it does: This paper proposes the use of diffusion models to generate aerial images with vehicle annotations and utilizes these synthetic data to train vehicle detectors, thereby achieving weakly supervised adaptation to unseen domains.
Adaptive Articulated Object Manipulation On The Fly with Foundation Model Reasoning and Part Grounding
Xiaojie Zhang (Beijing University of Posts and Telecommunications), Zhaofeng He (Beijing University of Posts and Telecommunications)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: The AdaRPG framework is proposed, which achieves component localization and segmentation, component suitability prediction through a base model, and generates high-level control code using LLM to enable cross-category adaptive manipulation of joint objects.
Adaptive Caching for Faster Video Generation with Diffusion Transformers
Kumara Kahatapitiya (Meta AI), Tian Xie (Meta AI)
GenerationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: This paper presents AdaCache, a training-free adaptive caching mechanism that accelerates the inference process of video diffusion Transformers by caching Transformer residuals and dynamically scheduling the cache rate based on content.
Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts
Zixuan Hu (Peking University), Ling-Yu Duan (Peking University)
Object DetectionDomain AdaptationAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: This paper proposes the DUO framework, which utilizes Test-Time Adaptation (TTA) technology to simultaneously minimize semantic uncertainty and geometric uncertainty in monocular 3D object detection, thereby enhancing detection performance in domain shift environments.
Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections
Youwei Zhou (Jiangnan University), Josef Kittler (University of Surrey)
RecognitionGraph Neural NetworkVideo
🎯 What it does: An adaptive hypergraph convolutional network (Hyper-GCN) is proposed, which constructs variable multi-vertex hypergraphs and introduces virtual connections to achieve efficient feature aggregation for skeleton sequences, thereby completing human action recognition based on skeletons.
Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation
Tao Lei (Shaanxi University of Science and Technology), Asoke K. Nandi (Brunel University of London)
SegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A semi-supervised medical image segmentation framework called ALHVR based on high-value region adaptive learning is proposed. It utilizes a dual-branch network to predict differences and classify pixels into three categories: reliable stable, reliable unstable, and unreliable stable regions, and designs an adaptive learning strategy specifically for high-value regions.
Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection
Yongkang Zhang (Beihang University), Zhong Zhou (Beihang University)
Object DetectionAnomaly DetectionTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: Utilizing adaptive prompt learning and Gaussian anomaly synthesis techniques to achieve out-of-distribution detection (OOD)
Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model
Zewei Xin (Shanghai Jiao Tong University), Guihai Chen (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerPrompt EngineeringMixture of ExpertsImageText
🎯 What it does: Proposes the RouteT2I framework, which dynamically selects between cloud-based large models or edge lightweight models for text-to-image generation, using multi-dimensional quality assessment and Pareto superiority for decision-making.
AdaptiveAE: An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes
Tianyi Xu (Shanghai AI Laboratory), Yujin Wang (Shanghai AI Laboratory)
RestorationConvolutional Neural NetworkReinforcement LearningOptical FlowImageVideo
🎯 What it does: Proposes AdaptiveAE, which uses reinforcement learning to adaptively control ISO and shutter speed, achieving high-quality HDR capture in dynamic scenes.
ADCD-Net: Robust Document Image Forgery Localization via Adaptive DCT Feature and Hierarchical Content Disentanglement
Kahim Wong (University of Macau), Jiantao Zhou (University of Electronic Science and Technology of China)
Object DetectionSegmentationConvolutional Neural NetworkMixture of ExpertsContrastive LearningImage
🎯 What it does: A network named ADCD-Net is proposed for precise localization of tampered areas in document images.
Adding Additional Control to One-Step Diffusion with Joint Distribution Matching
Yihong Luo (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
GenerationKnowledge DistillationDiffusion modelImage
🎯 What it does: This paper proposes a new method for incorporating unknown control signals into single-step diffusion generation—Joint Distribution Matching (JDM). It achieves the learning of unknown control from the teacher model by minimizing the upper bound of the inverse KL divergence of the image-conditional joint distribution.
Addressing Representation Collapse in Vector Quantized Models with One Linear Layer
Yongxin Zhu (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)
GenerationRepresentation LearningAuto EncoderImageAudio
🎯 What it does: By introducing a single-layer linear transformation reparameterized codebook into the vector quantization model, the issue of representation collapse in traditional VQ training is addressed.
Addressing Text Embedding Leakage in Diffusion-based Image Editing
Sunung Mun (POSTECH), Jungseul Ok (POSTECH)
Image TranslationGenerationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes a multi-target text editing framework called ALE based on diffusion models, aimed at eliminating attribute leakage issues and providing precise image editing without cross-target influence.
ADIEE: Automatic Dataset Creation and Scorer for Instruction-Guided Image Editing Evaluation
Sherry X. Chen (University of California), Suren Kumar (Samsung Electronics)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: An automated dataset generation and evaluation model, ADIEE, is proposed to assess the effects of instruction-driven image editing.
AdsQA: Towards Advertisement Video Understanding
Xinwei Long (Tsinghua University), Bowen Zhou (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningVideoMultimodalityBenchmark
🎯 What it does: Designed and constructed the AdsQA advertising video question-answering benchmark, and proposed the ReAd-R model based on reinforcement learning to evaluate the implicit reasoning ability of LLMs in advertising videos.
Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation
Yu Lei (Shanghai Jiao Tong University), Zhijie Deng (OPPO AI Center)
GenerationData SynthesisDiffusion modelScore-based ModelNeural Radiance FieldTextPoint Cloud
🎯 What it does: This paper proposes a text-to-3D generation method based on Linearized Lookahead Variational Score Distillation (L2-VSD). By improving the optimization order of the score model and the 3D model in VSD, it uses a linearized score model that retains only the first-order correction term to enhance generation quality.
Advancing Textual Prompt Learning with Anchored Attributes
Zheng Li (Nankai University), Jian Yang (Nankai University)
ClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A text prompting method based on general attribute anchoring (ATPrompt) is designed for visual-language models, enhancing the alignment capability between images and unknown categories by embedding attribute words into soft prompts.
Advancing Visual Large Language Model for Multi-granular Versatile Perception
Wentao Xiang (Tsinghua University), Yujiu Yang (Tsinghua University)
Object DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: A unified visual large language model framework MVP-LM is proposed, capable of simultaneously performing four types of visual perception tasks: box and mask detection, segmentation, and localization based on word/sentence instructions within the same model.
AdvDreamer Unveils: Are Vision-Language Models Truly Ready for Real-World 3D Variations?
Shouwei Ruan (Beihang University), Xingxing Wei (Tsinghua University)
GenerationPose EstimationAdversarial AttackTransformerVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: The AdvDreamer framework is proposed, which can generate adversarial 3D transformation (Adv-3DT) samples that can be reproduced in the real physical world using only single-view images, and thus evaluate the robustness of visual language models (VLM) under dynamic 3D changes.
Adversarial Attention Perturbations for Large Object Detection Transformers
Zachary Yahn (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
Object DetectionAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes an adversarial attack method AFOG (Attention-Focused Offensive Gradient) aimed at large-scale detection Transformers, capable of uniformly attacking both Transformers and traditional CNN detectors.
Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization
Zuyu Zhang (Chongqing University of Posts and Telecommunications), Xu Zhang (Chongqing University of Posts and Telecommunications)
Domain AdaptationOptimizationAdversarial AttackImage
🎯 What it does: Proposes the LEAwareSGD optimizer, which guides training close to the edge of chaos through the Lyapunov exponent, achieving data augmentation for single-domain generalization.
Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis
Yanzuo Lu (Sun Yat-Sen University), Jian-Huang Lai (Sun Yat-Sen University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationDiffusion modelGenerative Adversarial NetworkImageVideoTextOrdinary Differential Equation
🎯 What it does: Proposes a two-stage distillation method, ADM and ADP, for pre-trained diffusion models like SDXL, achieving a significant reduction in sampling steps.
Adversarial Exploitation of Data Diversity Improves Visual Localization
Sihang Li (New York University), Yiming Li (New York University)
Data SynthesisPose EstimationAutonomous DrivingAdversarial AttackTransformerGenerative Adversarial NetworkGaussian SplattingImage
🎯 What it does: This paper proposes a framework that utilizes 3D Gaussian splatting to synthesize diverse appearance data, combined with a two-branch joint training framework with an adversarial discriminator, to enhance the robustness of Absolute Pose Regression (APR) in visual localization.
Adversarial Purification via Super-Resolution and Diffusion
Mincheol Park (Samsung Advanced Institute of Technology), Suhyun Kim (Kyung Hee University)
ClassificationRestorationSuper ResolutionAdversarial AttackConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes a novel adversarial purification process that combines super-resolution (SR) with diffusion models, called PuriFlow. It first upsamples and then downsamples to recover the early features of disturbed images using SR, followed by denoising with a diffusion model, thereby enhancing classification accuracy and robustness.
Adversarial Reconstruction Feedback for Robust Fine-grained Generalization
Shijie Wang (Shandong University of Science and Technology), Haojie Li (Dalian University of Technology)
RetrievalKnowledge DistillationRepresentation LearningConvolutional Neural NetworkAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: The paper proposes an adversarial reconstruction feedback framework (AdvRF) to improve the differential representation in fine-grained image retrieval, making it category-independent.
Adversarial Robust Memory-Based Continual Learner
Xiaoyue Mi (Institute of Computing Technology, Chinese Academy of Sciences), Yang Liu (Institute for AI Industry Research, Tsinghua University)
Adversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: A memory-based continual learning framework that is robust to adversarial attacks is proposed, addressing the issues of rapid forgetting and gradient blurring caused by adversarial training in continual learning.
Adversarial Robustness of Discriminative Self-Supervised Learning in Vision
Ömer Veysel Çağatan (Koç University), M. Emre Gursoy
ClassificationObject DetectionSegmentationAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A systematic evaluation of the adversarial robustness of seven discriminative self-supervised learning models and one supervised model across multiple tasks such as ImageNet classification, transfer learning, segmentation, and detection.
Adversarial Training for Probabilistic Robustness
Yi Zhang (University of Warwick), Xingyu Zhao (University of Warwick)
ClassificationAdversarial AttackConvolutional Neural NetworkTransformerAuto EncoderImage
🎯 What it does: A method for adversarial training aimed at probabilistic robustness, called AT-PR, is proposed to enhance the model's probability (PR) of making incorrect classifications within a perturbation ball.
AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations
Junli Liu (Northwestern Polytechnical University), Bin Zhao (Northwestern Polytechnical University)
RecognitionObject DetectionTransformerImageTextBenchmark
🎯 What it does: This paper proposes the Aerial Visual Grounding (AerialVG) task, constructs the first aerial visual grounding dataset containing 5K high-resolution aerial images, 50K natural language descriptions, and 103K objects, and presents a dedicated model to achieve this task.
Aether: Geometric-Aware Unified World Modeling
Haoyi Zhu (USTC), Tong He (USTC)
GenerationDepth EstimationDomain AdaptationDiffusion modelAuto EncoderSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: AETHER is proposed, a unified 4D world model that integrates three main capabilities: 4D dynamic reconstruction, action-based future video prediction, and visual planning.
AffordDexGrasp: Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance
Yi-Lin Wei (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
OptimizationRobotic IntelligenceTransformerLarge Language ModelFlow-based ModelImageMultimodality
🎯 What it does: This paper proposes an open-ended language-guided multi-finger grasping framework based on intermediate affordance representation (AffordDexGrasp), which enables the understanding and execution of grasping instructions for unknown category objects.
After the Party: Navigating the Mapping From Color to Ambient Lighting
Florin-Alexandru Vasluianu (University of Wurzburg), Radu Timofte (University of Wurzburg)
Image TranslationRestorationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a method for ambient light normalization under a colorful lighting environment, called RLN 2, and constructs a large high-resolution dataset CL3AN for training and evaluation.
AFUNet: Cross-Iterative Alignment-Fusion Synergy for HDR Reconstruction via Deep Unfolding Paradigm
Xinyue Li (Tongji University), Wenhan Yang (Pengcheng Laboratory)
RestorationTransformerImage
🎯 What it does: This paper proposes a cross-iteration alignment-fusion deep unfolding network (AFUNet) for reconstructing high dynamic range (HDR) images from multi-exposure low dynamic range images.
AG2aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editing
Zhaonan Wang (Shandong University), Changhe Tu (Shandong University)
Object DetectionSegmentationRetrievalGraph Neural NetworkContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: Proposes an Anchor-Graph structured Gaussian splatting method that organizes semantic features with anchor points and regulates Gaussian primitives to achieve instance-level 3D scene understanding and editing.
AGO: Adaptive Grounding for Open World 3D Occupancy Prediction
Peizheng Li (Mercedes-Benz AG), Andreas Zell (University of Tübingen)
RecognitionObject DetectionSegmentationAutonomous DrivingTransformerVision Language ModelContrastive LearningImagePoint Cloud
🎯 What it does: Proposes the AGO framework, which is based on VLM knowledge self-supervised learning for 3D semantic occupancy prediction, and achieves recognition of known and unknown objects in open-world scenarios.
Agreement aware and dissimilarity oriented GLOM
Ru Zeng (University of Shanghai for Science and Technology), Hui Yu (University of Glasgow)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: An improved GLOM model is proposed, incorporating a contrastive consistency enhancer and a diversity focus head to enhance the formation of embedding islands and the computational efficiency of the model.