ICCV 2025 Papers — Page 2
IEEE/CVF International Conference on Computer Vision · 2701 papers
AgroBench: Vision-Language Model Benchmark in Agriculture
Risa Shinoda (Osaka University), Yoshitaka Ushiku (OMRON SINIC X)
ClassificationRecognitionTransformerVision Language ModelImageTextMultimodalityBenchmarkAgriculture RelatedChain-of-Thought
🎯 What it does: AgroBench has been developed, a comprehensive benchmark for agricultural visual language models (VLMs), covering 7 tasks (disease identification, pest identification, weed identification, crop management, disease management, machinery usage, traditional management), with a total of 4,342 QA pairs and approximately 4,218 high-quality farmland images constructed through expert annotations.
AHCPTQ: Accurate and Hardware-Compatible Post-Training Quantization for Segment Anything Model
Wenlun Zhang (Keio University), Kentaro Yoshioka (Keio University)
Object DetectionSegmentationComputational EfficiencyImage
🎯 What it does: A post-training quantization framework AHCPTQ is proposed for the Segment Anything Model (SAM), addressing the issues of activation distribution skew and excessive differences between channels in SAM quantization.
AIComposer: Any Style and Content Image Composition via Feature Integration
Haowen Li (Peking University), Yunjin Li (Beijing Yuanli Science and Technology Co., Ltd.)
GenerationData SynthesisDiffusion modelImageBenchmark
🎯 What it does: Proposes AIComposer, which achieves cross-domain image synthesis using linear/non-linear fusion of CLIP features without the need for text prompts;
AID: Adapting Image2Video Diffusion Models for Instruction-guided Video Prediction
Zhen Xing (Fudan University), Yu-Gang Jiang (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: This paper proposes the AID model, which targets the text-guided video prediction task. It utilizes a pre-trained Image2Video Diffusion model combined with a multimodal large language model (MLLM) to predict future frames of videos, and integrates multimodal conditions into UNet through a Dual Query Transformer (DQFormer). Additionally, it designs three types of adapters—spatial, short-term, and long-term—to achieve parameter-efficient transfer.
AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models
Ziyin Zhou (Xiamen University), Rongrong Ji (Xiamen University)
Object DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageMultimodality
🎯 What it does: An explainable and generalizable AI-generated image detection method called AIGI-Holmes based on a multimodal large language model is proposed, along with the construction of the Holmes-Set dataset, which provides human-verifiable explanations.
AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning
Yiwu Zhong (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoMultimodality
🎯 What it does: An adaptive reasoning method that is training-independent is proposed, which significantly reduces computational load by first merging visual tokens in a multimodal large language model and then performing token pruning at the LLM layer.
AIM: Amending Inherent Interpretability via Self-Supervised Masking
Eyad Alshami (Max Planck Institute for Informatics), Margret Keuper (Max Planck Institute for Informatics)
Explainability and InterpretabilityConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The AIM method is proposed, which encourages CNNs to focus on true features through a self-supervised masking mechanism, thereby enhancing interpretability and robustness against adversarial pseudo-features.
AIRA: Activation-Informed Low-Rank Adaptation for Large Models
Lujun Li (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyData-Centric LearningTransformerSupervised Fine-TuningDiffusion modelImageMultimodality
🎯 What it does: AIRA is proposed, a low-rank adaptation framework based on activation information, which improves the initialization, hierarchical rank allocation, and training process of LoRA to enhance the efficiency of fine-tuning large models.
AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference
Kai Huang (Alibaba Group), Hao Wang (Alibaba Group)
CompressionComputational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: This paper proposes a KV cache compression method called AirCache based on cross-modal correlation, aimed at significantly reducing cache usage and latency during inference while maintaining the performance of visual language models.
AJAHR: Amputated Joint Aware 3D Human Mesh Recovery
Hyunjin Cho (Chung-Ang University), Jongwon Choi (Chung-Ang University)
GenerationPose EstimationTransformerMultimodalityMesh
🎯 What it does: Proposed the AJAHR framework to achieve 3D human mesh recovery for amputee populations;
Align Your Rhythm: Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation
Congyi Fan (Harbin Engineering University), Haiwei Pan (Harbin Engineering University)
GenerationData SynthesisPose EstimationTransformerAuto EncoderVideoAudio
🎯 What it does: This work proposes the Danceba framework, which achieves automatic generation of 3D dance movements based on music, focusing on improving the precise alignment of dance with music rhythm, naturalness, and diversity.
AlignDiff: Learning Physically-Grounded Camera Alignment via Diffusion
Liuyue Xie (Carnegie Mellon University), Zhiheng Jia (Google)
Pose EstimationOptimizationTransformerDiffusion modelVideo
🎯 What it does: A diffusion-based camera calibration framework named AlignDiff is proposed, which can jointly estimate the camera's intrinsic parameters, extrinsic parameters, and optical distortion using ray bundle representation.
AlignGuard: Scalable Safety Alignment for Text-to-Image Generation
Runtao Liu (Hong Kong University of Science and Technology), Fabio Pizzati (University of Oxford)
GenerationOptimizationSafty and PrivacyDiffusion modelImageText
🎯 What it does: This paper proposes AlignGuard, which utilizes DPO and LoRA experts to securely align text-to-image models, removing up to 723 harmful concepts without affecting generation quality.
Aligning Constraint Generation with Design Intent in Parametric CAD
Evan Casey (Autodesk Research), Karl D.D. Willis (Autodesk Research)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
🎯 What it does: In CAD parametric design, 'design intent alignment' for constraint generation is achieved by integrating feedback from constraint solvers.
Aligning Effective Tokens with Video Anomaly in Large Language Models
Yingxian Chen (University of Hong Kong), Yik-Chung Wu (University of Hong Kong)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: A multi-modal large language model VA-GPT for video anomaly detection and summarization is proposed, which can efficiently locate abnormal events and generate text descriptions.
Aligning Global Semantics and Local Textures in Generative Video Enhancement
Zhikai Chen (University of Science and Technology of China), Tao Mei (HiDream.ai Inc.)
RestorationGenerationDiffusion modelVideo
🎯 What it does: The Generative Video Enhancement framework GenVE, based on diffusion models, utilizes high-quality image references to achieve global semantic and local texture alignment for low-quality videos, enhancing video details and visual quality.
Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching
Yang Liu (Sichuan University), Jiancheng Lv (Sichuan University)
RetrievalKnowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This study investigates the issue of information density imbalance in image-text matching and proposes a two-stage training framework D2S-VSE, which enhances the information capacity of visual embeddings through dense text pre-training and improves the information capacity of sparse text embeddings during the fine-tuning stage using dense-to-sparse feature distillation.
Aligning Moments in Time using Video Queries
Yogesh Kumar (Indian Institute of Technology Jodhpur), Anand Mishra (Microsoft)
RetrievalTransformerVision Language ModelVideo
🎯 What it does: The paper proposes a dual-stage sequence alignment model called MATR based on Transformer, which can accurately locate semantically matching segments in the target video based on video queries.
Aligning Vision to Language: Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning
Junming Liu (Tongji University), Botian Shi (New York University)
ClassificationRetrievalCompressionTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: Construct a multimodal knowledge graph (MMKG) without manual annotation, generating image descriptions through a vision-language model and integrating them with textual information, thereby enhancing the cross-modal reasoning ability of large language models (LLMs).
All in One: Visual-Description-Guided Unified Point Cloud Segmentation
Zongyan Han (Mohamed Bin Zayed University of Artificial Intelligence), Rao Muhammad Anwer (Mohamed Bin Zayed University of Artificial Intelligence)
SegmentationTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityPoint Cloud
🎯 What it does: This paper proposes the VDG-Uni3DSeg framework, which utilizes pre-trained vision-language models (CLIP) and category descriptions generated by large language models (LLM), along with reference images retrieved from the internet, to integrate multimodal queries into the 3D point cloud segmentation task, achieving a unified approach for semantic, instance, and panoptic segmentation.
All Parts Matter: A Unified Mask-Free Virtual Try-On Framework
Chenghu Du (Wuhan University of Technology), Yi Rong (Wuhan University of Technology)
GenerationData SynthesisLarge Language ModelDiffusion modelImage
🎯 What it does: A unified mask-free virtual try-on framework is proposed, which enhances generation quality through text-driven pseudo-sample preparation, gated fusion, texture-aware injection, and conditional reasoning fusion.
Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation
Shuchang Ye (University of Sydney), Jinman Kim (University of Sydney)
SegmentationVision Language ModelImageBiomedical Data
🎯 What it does: ProLearn framework is proposed, achieving medical image segmentation through prototype-driven semantic approximation without the need for text input during the inference phase.
AllGCD: Leveraging All Unlabeled Data for Generalized Category Discovery
Xinzi Cao (Sun Yat-sen University), Yutong Lu (Sun Yat-sen University)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: The AllGCD method is proposed, which significantly enhances the ability to distinguish between known and unknown categories by integrating all unlabeled data into supervised contrastive learning.
Allowing Oscillation Quantization: Overcoming Solution Space Limitation in Low Bit-Width Quantization
Weiying Xie (Xidian University), Yunsong Li (Xidian University)
OptimizationKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the Allowable Oscillation Quantization (AOQ) method, which encourages weight quantization threshold oscillation in the early training phase and suppresses oscillation in the later phase, thereby expanding the solution space for low-bit-width quantization and improving model accuracy.
AllTracker: Efficient Dense Point Tracking at High Resolution
Adam W. Harley (Stanford University), Leonidas Guibas (Stanford University)
Object TrackingComputational EfficiencyTransformerOptical FlowVideo
🎯 What it does: AllTracker achieves long-term tracking of all pixels by computing long-range optical flow between the query frame and all frames in the video in a single pass;
ALOcc: Adaptive Lifting-Based 3D Semantic Occupancy and Cost Volume-Based Flow Predictions
Dubing Chen (University of Macau), Jianbing Shen (University of Macau)
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkOptical FlowImagePoint Cloud
🎯 What it does: A vision-based framework is proposed for simultaneously predicting 3D semantic occupancy states and motion flow.
Always Skip Attention
Yiping Ji (Adelaide University), Simon Lucey (Adelaide University)
OptimizationComputational EfficiencyTransformerImage
🎯 What it does: This paper demonstrates that the self-attention blocks in visual Transformers have a fundamentally poor condition number without skip connections, leading to training difficulties, and proposes a preprocessing method called Token Greying, which involves singular value decomposition or discrete cosine transform on the input tokens to improve the condition number, enhancing model stability and performance.
AM-Adapter: Appearance Matching Adapter for Exemplar-based Semantic Image Synthesis in-the-Wild
Siyoon Jin (Korea Advanced Institute of Science and Technology), Seungryong Kim (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes AM-Adapter, which can achieve semantic image synthesis based on example images in outdoor scenes, preserving local appearance and semantic structure.
AMD: Adaptive Momentum and Decoupled Contrastive Learning Framework for Robust Long-Tail Trajectory Prediction
Bin Rao (University of Macau), Zhenning Li (University of Macau)
Autonomous DrivingOptimizationRecurrent Neural NetworkTransformerContrastive LearningTime SeriesSequential
🎯 What it does: An adaptive momentum and decoupled contrastive learning framework named AMD is proposed to address the long-tail distribution problem in traffic trajectory prediction.
AMDANet: Attention-Driven Multi-Perspective Discrepancy Alignment for RGB-Infrared Image Fusion and Segmentation
Haifeng Zhong (Jilin University), Yixing Gao (Chinese Academy of Sciences)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: Proposes AMDANet, an attention-driven multi-view disparity alignment network for RGB-infrared image fusion and semantic segmentation.
Amodal Depth Anything: Amodal Depth Estimation in the Wild
Zhenyu Li (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
Depth EstimationImage
🎯 What it does: This study investigates the prediction of the relative depth of occluded objects in real-world scenarios and proposes a method for amodal depth estimation in the wild.
Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images
Tianhao Wu (Nanyang Technological University), Tat-Jen Cham (Singapore Institute of Technology)
GenerationData SynthesisDiffusion modelImagePoint Cloud
🎯 What it does: A single-stage Amodal3R model is proposed, capable of directly generating complete 3D geometry and realistic textures from a single image with partial occlusion;
An Efficient Hybrid Vision Transformer for TinyML Applications
Fanhong Zeng (Xidian University), Rui Lai (Xidian University)
ClassificationObject DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: This paper presents TinyNeXt, an efficient hybrid Vision Transformer architecture designed for TinyML.
An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval
Jaeseok Byun (Seoul National University), Taesup Moon (Seoul National University)
RetrievalVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a post-processing framework RTD to reduce the task gap of text encoders in zero-shot synthesized image retrieval tasks.
An Empirical Study of Autoregressive Pre-training from Videos
Jathushan Rajasegaran (Meta FAIR), Jitendra Malik (University of California, Berkeley)
Object TrackingData SynthesisRepresentation LearningTransformerSupervised Fine-TuningImageVideo
🎯 What it does: This paper proposes and experiments with a video pre-training framework called Toto based on autoregressive transformers, utilizing discrete visual tokens to pre-train on large-scale unlabeled videos and performing transfer learning on various downstream tasks.
An Information-Theoretic Regularizer for Lossy Neural Image Compression
Yingwen Zhang (City University of Hong Kong), Shiqi Wang (Bytedance Inc)
CompressionDomain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: By incorporating conditional source entropy regularization into the training objective of the neural image compression network, the compression performance and cross-domain generalization ability have been improved.
An Inversion-based Measure of Memorization for Diffusion Models
Zhe Ma (Zhejiang University), Wenzhi Chen (Zhejiang University)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: A reverse-based memory measurement method called InvMM is proposed, which quantifies the degree of memory retention of a single image by utilizing the sensitive noise distribution of diffusion models.
An OpenMind for 3D Medical Vision Self-supervised Learning
Tassilo Wald (German Cancer Research Center), Klaus Maier-Hein
ClassificationSegmentationConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The largest publicly available 3D medical imaging pre-training dataset, OpenMind, has been proposed, and various self-supervised learning methods have been uniformly evaluated on it. The performance of CNN and Transformer architectures has been compared, and pre-training and fine-tuning frameworks and models have been released.
Analyzing Finetuning Representation Shift for Multimodal LLMs Steering
Pegah Khayatan (Sorbonne Université), Matthieu Cord (Valeo)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This study analyzes and controls concept transfer during the fine-tuning process of multimodal large language models, proposing a concept-level interpretability framework.
Anchor Token Matching: Implicit Structure Locking for Training-free AR Image Editing
Taihang Hu (Nankai University), Ming-Ming Cheng (Nankai University)
Image TranslationGenerationTransformerAuto EncoderImage
🎯 What it does: This paper proposes a training-free, zero-shot autoregressive image editing method called ISLock, which utilizes Anchor Token Matching (ATM) to implicitly lock the structure during the decoding process, enabling tasks such as object replacement, addition, removal, attribute modification, and style transfer.
AnimalClue: Recognizing Animals by their Traces
Risa Shinoda (University of Osaka), Hirokatsu Kataoka (Visual Geometry Group, University of Oxford)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: Constructed the AnimalClue dataset, covering five types of animal traces: footprints, feces, eggs, bones, and feathers, and providing four benchmarks for classification, detection, instance segmentation, and ecological feature prediction.
Animate Anyone 2: High-Fidelity Character Image Animation with Environment Affordance
Li Hu (Alibaba Group), Liefeng Bo (Alibaba Group)
GenerationData SynthesisDiffusion modelVideoBenchmark
🎯 What it does: A role animation framework called Animate Anyone 2 based on diffusion models is proposed, which achieves realistic collaboration between characters and the environment while maintaining consistency in character actions.
AnimateAnyMesh: A Feed-Forward 4D Foundation Model for Text-Driven Universal Mesh Animation
Zijie Wu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
GenerationTransformerRectified FlowAuto EncoderTextMesh
🎯 What it does: An end-to-end Feed-Forward 4D foundational model, AnimateAnyMesh, is proposed, capable of generating high-quality animations for any 3D mesh based on text prompts.
AnimeGamer: Infinite Anime Life Simulation with Next Game State Prediction
Junhao Cheng (City University of Hong Kong), Ying Shan
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoMultimodality
🎯 What it does: A multimodal large model-based infinite anime life simulation system called AnimeGamer has been constructed, which can generate continuous multi-round game states based on open player instructions, including dynamic animation clips and character attribute updates.
AnnofreeOD: Detecting All Classes at Low Frame Rates Without Human Annotations
Boyi Sun (Institute of Automation, Chinese Academy of Sciences), Fei-Yue Wang (Institute of Automation, Chinese Academy of Sciences)
Object DetectionAutonomous DrivingKnowledge DistillationPoint Cloud
🎯 What it does: A framework for 3D object detection without manual annotation, called AnnofreeOD, has been developed, which generates pseudo 3D boxes based on 2D to 3D knowledge distillation and uses noise-robust regression.
Anomaly Detection of Integrated Circuits Package Substrates Using the Large Vision Model SAIC: Dataset Construction, Methodology, and Application
Ruiyun Yu (Northeastern University), Haoyuan Li (Northeastern University)
SegmentationAnomaly DetectionKnowledge DistillationTransformerPrompt EngineeringImage
🎯 What it does: Designed and released the largest-scale integrated circuit ceramic substrate surface defect 2D anomaly detection dataset CPS2D-AD, and proposed the embedded knowledge distillation model SAIC based on SAM for efficient and accurate defect segmentation and localization.
Anti-Tamper Protection for Unauthorized Individual Image Generation
Zelin Li (University of Illinois Urbana-Champaign), Dong Wang (University of Illinois Urbana-Champaign)
GenerationData SynthesisSafty and PrivacyAdversarial AttackConvolutional Neural NetworkDiffusion modelAuto EncoderImage
🎯 What it does: A tamper-resistant perturbation (ATP) mechanism is proposed, which can inject protective perturbations into images to resist personalized image generation, and can detect when perturbations are tampered with during purification attacks, thereby rejecting illegal generation requests.
Any-SSR: How Recursive Least Squares Works in Continual Learning of Large Language Model
Kai Tong (South China University of Technology), Huiping Zhuang (South China University of Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The Analytic Subspace Routing (Any-SSR) framework is proposed to achieve continuous learning of large language models without using historical data. It avoids catastrophic forgetting by freezing low-level features, training independent LoRA subspaces for each task, and using a recursive least squares algorithm to construct task routers.
Any2AnyTryon: Leveraging Adaptive Position Embeddings for Versatile Virtual Clothing Tasks
Hailong Guo (Beijing University of Posts and Telecommunications), Chuang Zhang (Beijing University of Posts and Telecommunications)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodality
🎯 What it does: A novel, mask-free virtual try-on framework called Any2AnyTryon is proposed, capable of simultaneously performing multi-task generation such as clothing reconstruction, model-free try-on, and layered try-on based on text descriptions and model/clothing images.
AnyBimanual: Transferring Unimanual Policy for General Bimanual Manipulation
Guanxing Lu (Tsinghua University), Ziwei Wang (Nanyang Technological University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelMultimodality
🎯 What it does: A Plug-and-Play framework called AnyBimanual is proposed, which transfers pre-trained single-arm policies to dual-arm operations, utilizing a skill manager to dynamically schedule shared skills and a visual aligner to mitigate observation distribution differences.
AnyCalib: On-Manifold Learning for Model-Agnostic Single-View Camera Calibration
Javier Tirado-Garín (University of Zaragoza), Javier Civera (University of Zaragoza)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: A single-image camera calibration method named AnyCalib is proposed, supporting perspective, distortion, and editing (cropping/stretching) of images, and is independent of the camera model.
AnyI2V: Animating Any Conditional Image with Motion Control
Ziye Li (Fudan University), Henghui Ding (Fudan University)
GenerationData SynthesisDiffusion modelVideoMultimodalityPoint Cloud
🎯 What it does: A training-free image-to-video (I2V) generation framework called AnyI2V is proposed, which can utilize the first frame conditional image of any modality and user-defined motion trajectories to control video content and motion.
AnyPortal: Zero-Shot Consistent Video Background Replacement
Wenshuo Gao (Peking University), Shuai Yang (Peking University)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: The zero-shot ANYPORTAL framework is proposed, achieving video background replacement and foreground relighting;
AR-1-to-3: Single Image to Consistent 3D Object via Next-View Prediction
Xuying Zhang (Nankai University), Ming-Ming Cheng (Nankai University)
GenerationData SynthesisDiffusion modelAuto EncoderImagePoint Cloud
🎯 What it does: A framework called AR-1-to-3 is proposed for generating consistent 3D objects from single-view images, utilizing a diffusion model to progressively predict new views in order of distance and generate complete multi-view images.
AR-VRM: Imitating Human Motions for Visual Robot Manipulation with Analogical Reasoning
Dejie Yang (Peking University), Yang Liu (Peking University)
Pose EstimationRobotic IntelligenceTransformerVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes AR-VRM, which utilizes large-scale human hand keypoint video to learn explicit action knowledge and maps human hand actions to robot execution instructions through similar reasoning, thereby completing visual robotic manipulation tasks.
ArchiSet: Benchmarking Editable and Consistent Single-View 3D Reconstruction of Buildings with Specific Window-to-Wall Ratios
Jun Yin (Tsinghua University), Shuai Lu (Tsinghua University)
GenerationTransformerDiffusion modelNeural Radiance FieldImageMeshBenchmark
🎯 What it does: This paper proposes a new task in the field of architectural design—generating editable and proportionally consistent 3D architectural models from single-view images, and constructs a novel 3D architectural dataset called ArchiSet. Subsequently, a BuildingMesh framework based on diffusion models is designed to achieve this task.
Are They the Same? Exploring Visual Correspondence Shortcomings of Multimodal LLMs
Yikang Zhou (Wuhan University), Xiangtai Li (Bytedance Seed)
RecognitionObject DetectionObject TrackingSegmentationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningVideoTextMultimodalityBenchmark
🎯 What it does: A benchmark called MMVM was proposed to evaluate the visual correspondence capability of multimodal large language models, and a multiple-choice visual correspondence SFT dataset containing 220K reasoning texts was constructed. Based on this dataset, the CoLVA model was trained, significantly improving performance on visual correspondence tasks.
Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data and Metric Perspectives
Shaoyuan Xie (University of California), Liang Pan (Shanghai AI Laboratory)
Object DetectionAutonomous DrivingExplainability and InterpretabilityTransformerVision Language ModelImageTextBenchmark
🎯 What it does: This paper proposes the DriveBench benchmark, which systematically evaluates the reliability and interpretability of 12 VLMs under 17 types of visual distortions and 4 major driving tasks using 19,200 frames of images and 20,498 question-answer pairs. It finds that VLMs often generate seemingly reasonable answers based on general knowledge or textual cues, lacking true reliance on visual input. Experimental validation reveals that data bias and insufficient evaluation metrics lead to misjudgments of robustness. Furthermore, a RAU framework is constructed, utilizing VLM's perception of distortion types and tool-driven denoising to enhance the robustness of 3D detection under degraded images in RoboBEV. The contributions of this paper include providing robust evaluations across multiple scenarios, revealing the foundational visual defects and metric limitations of VLMs, and proposing a proxy-based robust perception method, although it is still limited by the coverage of the dataset and the generalization of tools.
ArgMatch: Adaptive Refinement Gathering for Efficient Dense Matching
Yuxin Deng (Wuhan University), Jiayi Ma (Wuhan University)
Image TranslationObject DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkOptical FlowImage
🎯 What it does: An adaptive refinement collection (ArgMatch) process is proposed in panoramic geometric matching, utilizing lightweight feature extraction, global matching, and a three-stage refinement (offset estimator, local consistency corrector, local consistency upsampler) to achieve efficient dense correspondence.
ArgoTweak: Towards Self-Updating HD Maps through Structured Priors
Lena Wild (KTH Royal Institute of Technology), Patric Jensfelt (TRATON)
Autonomous DrivingExplainability and InterpretabilityTransformerPoint Cloud
🎯 What it does: A new dataset called ArgoTweak has been constructed, which is manually annotated and contains real map priors, sensor data, and updated maps. A bijective element-level change mapping framework has been proposed to achieve interpretability for HD map self-updating.
ARGUS: Hallucination and Omission Evaluation in Video-LLMs
Ruchit Rawal (University of Maryland), Tom Goldstein (University of Maryland)
Large Language ModelVideoTextBenchmark
🎯 What it does: The ARGUS framework is proposed to evaluate the hallucinations and omissions of large video models in free-text video subtitles, providing dual metrics (ArgusCost-H and ArgusCost-O).
ARIG: Autoregressive Interactive Head Generation for Real-time Conversations
Ying Guo (Meituan), Xiaoming Wei (Meituan)
GenerationData SynthesisDiffusion modelOptical FlowVideoMultimodality
🎯 What it does: This paper proposes a real-time interactive head generation framework based on continuous autoregression (AR) called ARIG, designed to generate high-quality and natural head movements in two-person dialogues in real-time.
ARMO: Autoregressive Rigging for Multi-Category Objects
Mingze Sun (Tsinghua University), Ruqi Huang (Tsinghua University)
GenerationData SynthesisPose EstimationDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: The ARMO framework is proposed to achieve automatic skeleton prediction and binding of 3D models based on autoregressive models and latent diffusion models.
ART: Adaptive Relation Tuning for Generalized Relation Prediction
Gopika Sudhakaran (TU Darmstadt), Stefan Roth (TU Darmstadt)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the Adaptive Relation Tuning (ART) framework for visual relationship detection (VRD), which fine-tunes visual language models (VLM) for relationship classification through instruction tuning and adaptive sampling, enhancing the model's reasoning and generalization capabilities on unseen relationships.
ArtEditor: Learning Customized Instructional Image Editor from Few-Shot Examples
Shijie Huang (National University of Singapore), Jiaming Liu (Alibaba Group)
Image TranslationGenerationTransformerDiffusion modelImage
🎯 What it does: This paper presents ArtEditor, an end-to-end image editing framework based on Diffusion Transformer, capable of learning from only a few examples and performing instruction-driven edits for specific styles.
Arti-PG: A Toolbox for Procedurally Synthesizing Large-Scale and Diverse Articulated Objects with Rich Annotations
Jianhua Sun (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
Data SynthesisPose EstimationPoint CloudMesh
🎯 What it does: This work proposes the Arti-PG toolbox, which can automatically synthesize thousands of diverse 3D articulated objects based on procedural structure descriptions and point correspondence techniques, providing rich annotations.
Articulate3D: Holistic Understanding of 3D Scenes as Universal Scene Description
Anna-Maria Halacheva (Sofia University St. Kliment Ohridski), Danda Pani Paudel (Sofia University St. Kliment Ohridski)
SegmentationPose EstimationTransformerPoint Cloud
🎯 What it does: This paper constructs the Articulate3D dataset and proposes the USDNet method to achieve segmentation of fine-grained components in indoor scenes from 3D point clouds and prediction of joint motion parameters.
ASCENT: Annotation-free Self-supervised Contrastive Embeddings for 3D Neuron Tracking in Fluorescence Microscopy
Haejun Han (Georgia Institute of Technology), Hang Lu (Georgia Institute of Technology)
Object TrackingRepresentation LearningTransformerContrastive LearningImageBiomedical Data
🎯 What it does: This paper presents ASCENT, a framework for 3D neuron tracking without manual annotation.
ASGS: Single-Domain Generalizable Open-Set Object Detection via Adaptive Subgraph Searching
Yuxuan Yuan (Xiamen University), Xinghao Ding (Xiamen University)
Object DetectionDomain AdaptationTransformerContrastive LearningImage
🎯 What it does: The ASGS framework is proposed for the single-source domain general open-set object detection task, combining subgraph unknown class learning and category embedding compression modules to achieve unknown class detection and cross-domain robustness.
Ask and Remember: A Questions-Only Replay Strategy for Continual Visual Question Answering
Imad Eddine Marouf (Institut Polytechnique de Paris), Joost Van De Weijer
Safty and PrivacyKnowledge DistillationTransformerImageText
🎯 What it does: Proposes a method for memory replay that only uses past task questions and achieves continual learning in visual question answering through attention consistency distillation.
AstroLoc: Robust Space to Ground Image Localizer
Gabriele Berton (Politecnico di Torino), Carlo Masone (Politecnico di Torino)
RetrievalContrastive LearningImage
🎯 What it does: Developed an image retrieval model called AstroLoc that can directly utilize astronaut aerial images for training, achieving precise localization from space to ground images.
Asynchronous Event Error-Minimizing Noise for Safeguarding Event Dataset
Ruofei Wang (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
Anomaly DetectionSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTime Series
🎯 What it does: A method for generating Unlearnable Examples for Asynchronous Event Streams (UEVs) is proposed, which ensures that event datasets are not used without authorization by constructing a noise suppression model that minimizes errors.
ATAS: Any-to-Any Self-Distillation for Enhanced Open-Vocabulary Dense Prediction
Juan Yeo (Seoul National University), Taesup Kim (Seoul National University)
Object DetectionSegmentationKnowledge DistillationContrastive LearningImage
🎯 What it does: This paper proposes the ATAS (Any-to-Any Self-Distillation) framework, which enhances the semantic consistency and fine-grained visual-language alignment of CLIP on unsupervised images through a global-local-global triple distillation, significantly improving open vocabulary dense prediction performance.
ATCTrack: Aligning Target-Context Cues with Dynamic Target States for Robust Vision-Language Tracking
Xiaokun Feng (University of Chinese Academy of Sciences), Kaiqi Huang (University of Chinese Academy of Sciences)
Object TrackingTransformerLarge Language ModelImageVideoTextMultimodality
🎯 What it does: Proposes the ATCTrack tracker, which enhances the robustness of visual-language tracking by utilizing dynamically aligned multimodal target-context information.
ATLAS: Decoupling Skeletal and Shape Parameters for Expressive Parametric Human Modeling
Jinhyung Park (Carnegie Mellon University), Rawal Khirodkar (Carnegie Mellon University)
GenerationPose EstimationOptimizationAuto EncoderMesh
🎯 What it does: The ATLAS model is proposed, achieving decoupling of human shape and skeletal parameters, providing high-resolution models, sparse nonlinear pose correction, and implementing a single-image fitting pipeline.
Att-Adapter: A Robust and Precise Domain-Specific Multi-Attributes T2I Diffusion Adapter via Conditional Variational Autoencoder
Wonwoong Cho (Purdue University), Yanxia Zhang (Toyota Research Institute)
Image TranslationGenerationDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes Att-Adapter, a pluggable module that enables continuous and multi-attribute fine control over pre-trained text-to-image diffusion models without the need for paired data.
Attention to Neural Plagiarism: Diffusion Models Can Plagiarize Your Copyrighted Images!
Zihang Zou (University of Central Florida), Liqiang Wang (University of Central Florida)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: This paper proposes a general neural plagiarism attack framework based on diffusion models, capable of copying copyrighted images without training or fine-tuning the model, and achieving forgery and ambiguity attacks in copyright protection systems by damaging or replacing watermarks (both visible and invisible).
Attention to the Burstiness in Visual Prompt Tuning!
Yuzhu Wang (Zhejiang Lab), Shu Kong
ClassificationObject DetectionSegmentationTransformerPrompt EngineeringImage
🎯 What it does: This paper studies the 'explosive' phenomenon in Visual Prompt Tuning (VPT) and proposes improvements to prompt learning through whitening and bilinear models, significantly enhancing the performance of pre-trained visual Transformers on downstream tasks.
Attention to Trajectory: Trajectory-Aware Open-Vocabulary Tracking
Yunhao Li (Institute of Software Chinese Academy of Sciences), Libo Zhang (Institute of Software Chinese Academy of Sciences)
Object TrackingTransformerLarge Language ModelContrastive LearningVideo
🎯 What it does: This paper proposes TRACT, an open vocabulary multi-object tracker that utilizes Trajectory Consistency Reinforcement (TCR) and Trajectory Enhanced Classification (TraCLIP) for trajectory awareness.
AU-Blendshape for Fine-grained Stylized 3D Facial Expression Manipulation
Hao Li (Beihang University), Junjun Pan (Beihang University)
GenerationTransformerAuto EncoderImageMesh
🎯 What it does: This paper constructs the AUBlendSet dataset based on AU-Blendshape and the AUBlendNet model, achieving fine-grained, stylized 3D facial expression manipulation for any identity.
Audio-visual Controlled Video Diffusion with Masked Selective State Spaces Modeling for Natural Talking Head Generation
Fa-Ting Hong (Hong Kong University of Science and Technology), Dan Xu
GenerationData SynthesisDiffusion modelVideoMultimodalityAudio
🎯 What it does: An end-to-end audio-visual controlled video diffusion framework called ACTalker is proposed, capable of generating natural speaker videos using audio and facial motion signals either simultaneously or individually.
Augmented and Softened Matching for Unsupervised Visible-Infrared Person Re-Identification
Zhiqi Pang (Harbin Institute of Technology), Junjie Wang (Nanjing Medical University)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an unsupervised visible-infrared person re-identification method called ASM, which enhances the robustness of pseudo-labels through cross-modal augmented matching and soft label momentum updates.
Augmented Mass-Spring Model for Real-Time Dense Hair Simulation
J. Alejandro Amador H. (King Abdullah University of Science and Technology), Dominik L. Michels (King Abdullah University of Science and Technology)
🎯 What it does: This paper presents an Incremental Mass-Spring (AMS) model that can simulate the dynamic behavior of tens of thousands of densely packed strands in real-time conditions.
Augmenting Moment Retrieval: Zero-Dependency Two-Stage Learning
Zhengxuan Wei (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
RetrievalKnowledge DistillationTransformerContrastive LearningVideo
🎯 What it does: An external dependency-free enhanced moment retrieval framework AMR is proposed, which improves moment localization accuracy through video splicing + enhancement strategies and two-stage training.
AURELIA: Test-time Reasoning Distillation in Audio-Visual LLMs
Sanjoy Chowdhury (University of Maryland), Dinesh Manocha (University of Maryland)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIVideoMultimodalityBenchmarkAudio
🎯 What it does: Proposed the AURELIA framework, which injects stepwise reasoning into inference through multi-agent interaction and constructs the AVReasonBench evaluation benchmark.
Authentic 4D Driving Simulation with a Video Generation Model
Lening Wang (Beihang University), Shanghang Zhang (Peking University)
GenerationData SynthesisDepth EstimationAutonomous DrivingDiffusion modelVideoPoint Cloud
🎯 What it does: A controllable 4D autonomous driving simulation framework called Stag is proposed, which reconstructs sparse 4D point clouds from panoramic images of real vehicles. It then generates continuous and realistic driving scenes from any perspective and time point through spatial-temporal separation of keyframe projection and a diffusion video generation model.
Auto-Controlled Image Perception in MLLMs via Visual Perception Tokens
Runpeng Yu (National University of Singapore), Xinchao Wang (National University of Singapore)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: By adding special tokens that can trigger visual perception to multimodal large language models, the model can autonomously control the visual encoding process.
Auto-Regressive Transformation for Image Alignment
Kanggeon Lee (Seoul National University), Kyoung Mu Lee (Kookmin University)
Image TranslationPose EstimationConvolutional Neural NetworkImageMagnetic Resonance Imaging
🎯 What it does: Proposes the Auto-Regressive Transformation (ART) method, which utilizes multi-scale features and autoregressive iterative refinement to achieve robust image alignment.
Auto-Regressively Generating Multi-View Consistent Images
JiaKui Hu (Peking University), Yanye Lu (Peking University)
GenerationData SynthesisTransformerLarge Language ModelImageMultimodality
🎯 What it does: A Multi-View Autoregressive (MV-AR) model is proposed to progressively generate multi-view consistent images under multimodal conditions such as text, images, camera poses, and geometric shapes.
Auto-Vocabulary Semantic Segmentation
Osman Ülger (University of Amsterdam), Martin R. Oswald (University of Amsterdam)
SegmentationTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: The AutoSeg method is proposed to automatically generate a vocabulary for images and perform semantic segmentation.
AutoComPose: Automatic Generation of Pose Transition Descriptions for Composed Pose Retrieval Using Multimodal LLMs
Yi-Ting Shen (University of Maryland), Shuvra S. Bhattacharyya (University of Maryland)
Pose EstimationRetrievalTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Utilizing a multimodal large language model to automatically generate pose transfer descriptions, and training a pose retrieval model based on this to address the cumbersome and inconsistent issues of traditional manual annotation.
Automated Model Evaluation for Object Detection via Prediction Consistency and Reliability
Seungju Yoo (Yonsei University), Kibok Lee (Yonsei University)
Object DetectionAutonomous DrivingImage
🎯 What it does: This paper proposes an unsupervised object detection model evaluation method called PCR (Prediction Consistency and Reliability), which estimates mAP by utilizing the spatial consistency and confidence information of candidate boxes generated by the detector before and after NMS.
Automated Red Teaming for Text-to-Image Models through Feedback-Guided Prompt Iteration with Vision-Language Models
Wei Xu (Wuhan University), Lina Wang (Wuhan University)
GenerationAdversarial AttackTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Designed and implemented an automatic red team framework FGPI based on a visual language model, used for iteratively optimizing aggressive prompts for text-to-image models, capable of inducing the generation of non-compliant images without leaking harmful text.
AutoOcc: Automatic Open-Ended Semantic Occupancy Annotation via Vision-Language Guided Gaussian Splatting
Xiaoyu Zhou (Wangxuan Institute of Computer Technology Peking University), Ming-Hsuan Yang (University of California Merced)
Object DetectionSegmentationAutonomous DrivingVision Language ModelGaussian SplattingImagePoint Cloud
🎯 What it does: We propose a fully automated, unsupervised 3D semantic occupancy annotation framework called AutoOcc, which can map multi-view images to dense occupancy voxels without relying on manual labels or predefined categories.
AutoPrompt: Automated Red-Teaming of Text-to-Image Models via LLM-Driven Adversarial Prompts
Yufan Liu (Institute of Information Engineering, Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering, Chinese Academy of Sciences)
GenerationSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: A black-box automatic red team framework called AutoPrompt based on LLM is proposed to generate readable adversarial suffixes that entice secure T2I models to produce inappropriate images.
Autoregressive Denoising Score Matching is a Good Video Anomaly Detector
Hanwen Zhang (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Anomaly DetectionTransformerDiffusion modelScore-based ModelVideo
🎯 What it does: A self-regressive denoising score matching (ADSM) framework is proposed, which uses a noise conditional score transformer to estimate the Stein score of videos and integrates scene, motion, and appearance information to achieve video anomaly detection.
AutoScape: Geometry-Consistent Long-Horizon Scene Generation
Jiacheng Chen (Simon Fraser University), Manmohan Chandraker (NEC Labs America)
GenerationData SynthesisAutonomous DrivingDiffusion modelImageVideoPoint Cloud
🎯 What it does: AutoScape proposes a hierarchical long-term driving scene generation framework, which first uses an RGB-D diffusion model to iteratively generate a few geometrically consistent keyframes, and then uses a video diffusion model to interpolate dense frames between keyframes, achieving high-quality, 3D-consistent driving videos lasting over 20 seconds.
Auxiliary Prompt Tuning of Vision-Language Models for Few-Shot Out-of-Distribution Detection
Wenjun Miao (Beihang University), Xiao Bai (Beihang University)
Anomaly DetectionTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Proposes the Auxiliary Prompt Tuning (APT) framework, which enhances CLIP-based few-shot OOD detection using pseudo OOD samples from external auxiliary data.
AV-Flow: Transforming Text to Audio-Visual Human-like Interactions
Aggelina Chatziagapi, Alexander Richard
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelVideoTextMultimodalityAudio
🎯 What it does: This paper presents AV-Flow, an end-to-end model that can generate photorealistic 4D talking avatars solely from text, capable of synchronously synthesizing speech, facial expressions, and head movements, and supporting always-online bidirectional interaction.
AV-Link: Temporally-Aligned Diffusion Features for Cross-Modal Audio-Video Generation
Moayed Haji-Ali (Rice University), Sergey Tulyakov (Snap Inc)
GenerationData SynthesisDiffusion modelVideoMultimodalityAudio
🎯 What it does: A unified framework called AV-Link is proposed, achieving cross-modal generation from video to audio (V2A) and from audio to video (A2V).