CVPR 2025 Papers with AI Summaries
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
→ CVPR 2025 papers with code (851)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 2871 CVPR 2025 papers by keyword, author or institution —
free trial on arXivSub.
2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification
Jingwei Zhang (Stony Brook University), Mahdi S. Hosseini (Concordia University)
ClassificationRepresentation LearningImage
🎯 What it does: This paper proposes 2DMamba, which utilizes a 2D selective state space model for efficient modeling of large-scale images and achieves spatial continuity in multi-instance learning.
3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes
Jan Held (University of Liège), Marc Van Droogenbroeck (University of Liège)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a rendering method based on 3D smooth convex bodies (3D Convex Splatting) for high-quality view synthesis, combining the advantages of real-time rendering speed and a small number of primitives;
3D Dental Model Segmentation with Geometrical Boundary Preserving
Shufan Xi (Beihang University), Aimin Hao (Southwest University)
SegmentationTransformerPoint CloudMesh
🎯 What it does: A cross-modal tooth segmentation method called CrossTooth is proposed, which utilizes curvature-based selective downsampling to preserve the gum boundary and integrates multi-view rendered image features to achieve high-precision tooth segmentation of 3D oral scan models.
3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations
Yating Wang (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
GenerationData SynthesisCompressionGaussian SplattingVideo
🎯 What it does: This paper constructs an animatable head avatar using 3D Gaussian Splatting and the FLAME facial model, employing tri-plane encoding for static neutral surface textures and lightweight encoding of dynamic textures for each expression deformation through 1D feature lines. It also introduces adaptive truncation opacity penalties and class-balanced sampling to enhance generalization to new expressions.
3D Gaussian Inpainting with Depth-Guided Cross-View Consistency
Sheng-Yu Huang (National Taiwan University), Yu-Chiang Frank Wang (NVIDIA)
RestorationGenerationDiffusion modelGaussian SplattingImage
🎯 What it does: A 3D scene filling framework based on 3D Gaussian Splatting (3DGIC) is proposed, achieving high-fidelity, residue-free object removal and filling from multiple viewpoints through depth-guided cross-view consistency.
3D Occupancy Prediction with Low-Resolution Queries via Prototype-aware View Transformation
Gyeongrok Oh (Korea University), Sangpil Kim (Korea University)
SegmentationAutonomous DrivingContrastive LearningPoint CloudBenchmark
🎯 What it does: A low-resolution query 3D occupancy prediction framework named ProtoOcc has been designed.
3D Prior Is All You Need: Cross-Task Few-shot 2D Gaze Estimation
Yihua Cheng (University of Birmingham), Hyung Jin Chang (Beihang University)
RecognitionDomain AdaptationTransformerImage
🎯 What it does: A cross-task few-shot 2D gaze estimation method is proposed, utilizing a pre-trained 3D gaze network and a differentiable projection module to achieve rapid adaptation of 2D gaze for unseen devices.
3D Student Splatting and Scooping
Jialin Zhu (University College London), He Wang (University College London)
GenerationData SynthesisOptimizationGaussian SplattingPoint Cloud
🎯 What it does: A Splatting and Scooping model based on student distribution is proposed, replacing the Gaussian in 3D Gaussian Splatting with negative components and a learnable tail thickness t-distribution to improve neural rendering.
3D-AVS: LiDAR-based 3D Auto-Vocabulary Segmentation
Weijie Wei (University of Amsterdam), Martin R. Oswald (University of Amsterdam)
Object DetectionSegmentationTransformerLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: Proposes the 3D-AVS method, which automatically generates a vocabulary for point clouds and performs semantic segmentation.
3D-GRAND: A Million-Scale Dataset for 3D-LLMs with Better Grounding and Less Hallucination
Jianing Yang, Joyce Chai
SegmentationGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextPoint CloudBenchmark
🎯 What it does: This work proposes a large-scale, dense ground 3D-text pairing dataset 3D-GRAND and a benchmark 3D-POPE for evaluating the hallucination behavior of 3D language models, and conducts large-scale instruction fine-tuning of 3D-LLM based on this.
3D-GSW: 3D Gaussian Splatting for Robust Watermarking
Youngdong Jang (Korea University), Sangpil Kim (Korea University)
Gaussian SplattingMesh
🎯 What it does: Design a robust watermark embedding method based on the 3D Gaussian Spray (3D-GS) model to achieve copyright protection for the model and its rendered images.
3D-HGS: 3D Half-Gaussian Splatting
Haolin Li (Northeastern University), Octavia Camps (Northeastern University)
GenerationOptimizationGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a 3D Half-Gaussian Splatting (3D-HGS) kernel as a pluggable improvement to 3D Gaussian Splatting (3D-GS) to enhance the rendering quality of high-frequency details and edges.
3D-LLaVA: Towards Generalist 3D LMMs with Omni Superpoint Transformer
Jiajun Deng (Australian Institute for Machine Learning), Ian Reid (Australian Institute for Machine Learning)
SegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningPoint Cloud
🎯 What it does: This paper presents 3D-LLaVA, a general-purpose 3D large language model that uses only point cloud input and is capable of dialogue, image-text interaction, and 3D point cloud segmentation.
3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning
Yuncong Yang (University of Massachusetts Amherst), Chuang Gan (Massachusetts Institute of Technology)
Robotic IntelligenceTransformerVision Language ModelPoint CloudBenchmark
🎯 What it does: A 3D scene memory framework called 3D-Mem is proposed, based on multi-view snapshots, for lifelong detection and reasoning.
3D-MVP: 3D Multiview Pretraining for Manipulation
Shengyi Qian (NVIDIA), Ankit Goyal (NVIDIA)
Robotic IntelligenceTransformerAuto EncoderPoint Cloud
🎯 What it does: Self-supervised pre-training of the multi-view Transformer of RVT is conducted through a multi-view masked autoencoder to learn spatial representations of 3D scenes, achieving a higher success rate in robotic grasping and manipulation tasks.
3D-SLNR: A Super Lightweight Neural Representation for Large-scale 3D Mapping
Chenhui Shi (Institute of Automation, Chinese Academy of Sciences), Yihong Wu (Institute of Automation, Chinese Academy of Sciences)
Autonomous DrivingComputational EfficiencyRepresentation LearningSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes 3D-SLNR, an ultra-lightweight neural representation that constructs a global SDF using local band-limited SDF through a small number of MLP parameters and learnable geometric attributes, achieving large-scale 3D mapping.
3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement
Yihang Luo (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
RestorationGenerationTransformerDiffusion modelImage
🎯 What it does: A 3DENHANCER framework based on a multi-view latent diffusion model has been developed to enhance the texture quality and cross-view consistency of low-quality multi-view images and their corresponding 3D models.
3DGUT: Enabling Distorted Cameras and Secondary Rays in Gaussian Splatting
Qi Wu (NVIDIA), Zan Gojcic (University of Toronto)
Autonomous DrivingComputational EfficiencyGaussian SplattingImage
🎯 What it does: This study investigates how to extend 3D Gaussian Splatting to support arbitrary nonlinear camera projections and secondary rays, achieving efficient real-time rendering through the Unscented Transform, compatible with rolling shutter, distorted cameras, and phenomena such as reflection and refraction.
3DTopia-XL: Scaling High-quality 3D Asset Generation via Primitive Diffusion
Zhaoxi Chen (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderMesh
🎯 What it does: This work presents 3DTopia-XL, a native 3D diffusion model that combines the atomic 3D representation PrimX with a diffusion Transformer, capable of efficiently generating high-quality GLB mesh assets directly usable for PBR rendering from text or images.
4D LangSplat: 4D Language Gaussian Splatting via Multimodal Large Language Models
Wanhua Li (Harvard University), Hanspeter Pfister (Harvard University)
SegmentationGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningGaussian SplattingVideoTextMultimodality
🎯 What it does: This paper proposes the 4D LangSplat method, which learns time-sensitive and time-independent language fields in the context of 4D Gaussian Splatting, enabling open vocabulary queries for dynamic scenes.
4D-Fly: Fast 4D Reconstruction from a Single Monocular Video
Diankun Wu (Tsinghua University), Yueqi Duan (Tencent)
Object TrackingSegmentationDepth EstimationOptimizationComputational EfficiencyGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: A streaming framework was constructed using explicit primitives based on 3D Gaussian splatting, capable of quickly reconstructing 4D scenes from monocular video in a matter of minutes and achieving long-distance point tracking.
4Deform: Neural Surface Deformation for Robust Shape Interpolation
Lu Sang (Technical University of Munich), Daniel Cremers (University of Bonn)
Auto EncoderPoint Cloud
🎯 What it does: This paper proposes a 4Deform method based on neural implicit surface deformation, which utilizes a velocity field and implicit field for continuous shape interpolation of sparse point cloud keyframes, supporting topological changes, non-isometric deformations, and local shapes.
4DGC: Rate-Aware 4D Gaussian Compression for Efficient Streamable Free-Viewpoint Video
Qiang Hu (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
CompressionGaussian SplattingVideo
🎯 What it does: A compression framework for free viewpoint video (FVV) based on 4D Gaussian compression is proposed, enabling variable bitrate streaming.
4DTAM: Non-Rigid Tracking and Mapping via Dynamic Surface Gaussians
Hidenobu Matsuki (Imperial College London), Andrew J. Davison (Imperial College London)
Object TrackingPose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingPoint CloudMesh
🎯 What it does: A 4D SLAM method called 4DTAM is proposed, which can jointly estimate camera pose, scene geometry, appearance, and non-rigid motion in real-time from a monocular RGB-D stream, achieving 4D reconstruction of dynamic scenes.
4Real-Video: Learning Generalizable Photo-Realistic 4D Video Diffusion
Chaoyang Wang (Snap Inc), Hsin-Ying Lee (Snap Inc)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Proposes the 4Real-Video framework, which can generate a complete 4D video mesh from fixed-view videos and frozen-time videos.
5%>100%: Breaking Performance Shackles of Full Fine-Tuning on Visual Recognition Tasks
Dongshuo Yin (Tsinghua University), Xue Yang (Shanghai Jiao Tong University)
RecognitionObject DetectionSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Multi-Cognitive Visual Adapter (Mona) tuning method for parameter-efficient fine-tuning of visual tasks while maintaining the advantages of pre-trained models.
A Bias-Free Training Paradigm for More General AI-generated Image Detection
Fabrizio Guillaro (University Federico II of Naples), Luisa Verdoliva (University Federico II of Naples)
Object DetectionTransformerDiffusion modelImage
🎯 What it does: A bias-free training paradigm (B-Free) is proposed, which generates counterfeit samples aligned with the semantics of real images through self-conditioned reconstruction and local inpainting, and trains a large-scale ViT model to detect AI-generated images.
A Closer Look at Time Steps is Worthy of Triple Speed-Up for Diffusion Model Training
Kai Wang (National University of Singapore), Yang You (National University of Singapore)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper presents SpeeD, a training acceleration method based on a detailed analysis of the time steps in diffusion models;
A Comprehensive Study of Decoder-Only LLMs for Text-to-Image Generation
Andrew Z. Wang (University of Washington), Yogesh Balaji (NVIDIA)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: The study uses decoder-only large language models (LLMs) as text encoders for text-to-image diffusion models and systematically evaluates the impact of different LLMs and their embedding extraction methods on image generation.
A Data-Centric Revisit of Pre-Trained Vision Models for Robot Learning
Xin Wen (Hong Kong University), Xiaojuan Qi (Hong Kong University)
Data-Centric LearningRobotic IntelligenceContrastive LearningImage
🎯 What it does: A systematic evaluation of the performance of visual models with different pre-training methods and data sources in robotic learning tasks (manipulation and perception) is conducted, and SlotMIM is proposed to learn more object-centric representations on non-single-object (NOC) data.
A Dataset for Semantic Segmentation in the Presence of Unknowns
Zakaria Laskar (Czech Technical University in Prague), C.V. Jawahar (Indian Institute of Information Technology Hyderabad)
SegmentationAnomaly DetectionImageTime SeriesBenchmark
🎯 What it does: A new semantic segmentation anomaly detection dataset, ISSU, is proposed, and various existing methods are benchmarked on this dataset.
A Distractor-Aware Memory for Visual Object Tracking with SAM2
Jovana Videnovic (University of Ljubljana), Matej Kristan (University of Ljubljana)
Object TrackingTransformerVideo
🎯 What it does: A novel 'distractor-aware memory' (DAM) for SAM2 has been designed and implemented, dividing memory into recent appearance memory (RAM) and distractor resolution memory (DRM). It proposes using multiple prediction information based on SAM2 outputs to detect key distractors and update DRM, significantly enhancing the robustness and accuracy of visual object tracking.
A Flag Decomposition for Hierarchical Datasets
Nathan Mankovich (Valencia University), Tolga Birdal (Imperial College London)
ClassificationRecognitionRestorationSegmentationSupervised Fine-TuningImage
🎯 What it does: This paper proposes Flag Decomposition (FD), a matrix decomposition method that preserves hierarchical structures, capable of mapping hierarchical data to the flag manifold for reconstruction, clustering, and few-shot learning.
A Focused Human Body Model for Accurate Anthropometric Measurements Extraction
Shuhang Chen (Fudan University), Shuigeng Zhou (Fudan University)
Pose EstimationDepth EstimationConvolutional Neural NetworkTransformerImagePoint CloudMesh
🎯 What it does: This paper proposes a focused human model that accurately extracts body measurements from a single image or Lidar point cloud by combining SMPLer-X, a Bypass network, and dynamic loss.
A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening
Jie Huang (University of Electronic Science and Technology of China), Liangjian Deng (University of Electronic Science and Technology of China)
RestorationSuper ResolutionImage
🎯 What it does: An Adaptive Dual-layer Weighting Mechanism (ADWM) is proposed, which adjusts feature heterogeneity and redundancy through Covariance-weighted (CACW) to achieve remote sensing panchromatic fusion.
A Hubness Perspective on Representation Learning for Graph-Based Multi-View Clustering
Zheming Xu (Beijing Jiaotong University), Michael C. Kampffmeyer (UiT The Arctic University of Norway)
Representation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A new framework called hubREP is proposed for graph-based multi-view clustering, aimed at addressing the hubness problem in high-dimensional embeddings to improve clustering performance.
A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions
Jiangbei Hu (Dalian University of Technology), Ying He (Nanyang Technological University)
RestorationSegmentationPoint Cloud
🎯 What it does: A lightweight UDF learning framework called LoSF-UDF based on local shape functions is proposed for reconstructing 3D surfaces from point clouds.
A New Statistical Model of Star Speckles for Learning to Detect and Characterize Exoplanets in Direct Imaging Observations
Théo Bodrito (École Normale Supérieure), Anne-Marie Lagrange (Université Grenoble Alpes)
Object DetectionSegmentationExplainability and InterpretabilityConvolutional Neural NetworkGaussian SplattingImageTime Series
🎯 What it does: An end-to-end method that integrates multi-scale statistical models with deep learning is proposed for the detection and photometric estimation of exoplanets in direct imaging.
A Physics-Informed Blur Learning Framework for Imaging Systems
Liqun Chen (Shanghai AI Laboratory), Tianfan Xue (The Chinese University of Hong Kong)
RestorationOptimizationTransformerSupervised Fine-TuningImagePhysics Related
🎯 What it does: A physics-based PSF learning framework is proposed, utilizing simple calibration steps and a two-stage learning process to estimate the spatially varying blur of the imaging system.
A Polarization-Aided Transformer for Image Deblurring via Motion Vector Decomposition
Duosheng Chen (Nankai University), Jufeng Yang (Nankai University)
RestorationTransformerOptical FlowImage
🎯 What it does: A motion decomposition Transformer (MDT) based on polar coordinates is proposed, which achieves deblurring by separating the translational and rotational motions of image blur.
A Regularization-Guided Equivariant Approach for Image Restoration
Yulu Bai (Xi'an Jiaotong University), Deyu Meng (Macau University of Science and Technology)
RestorationConvolutional Neural NetworkSupervised Fine-TuningImageComputed Tomography
🎯 What it does: A self-supervised regularization strategy (EQ-Reg) is proposed, which can achieve rotation equivariance in ordinary CNNs while maintaining high representation accuracy.
A Selective Re-learning Mechanism for Hyperspectral Fusion Imaging
Yuanye Liu (Hunan University), Shutao Li (Hunan University)
Image TranslationRestorationCompressionComputational EfficiencyTransformerImage
🎯 What it does: A selective re-learning hyperspectral fusion network, SRLF-Net, is proposed for the fusion of low-resolution hyperspectral images and multispectral images.
A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation
Zheng Zhang (Beijing University of Posts and Telecommunications), Wendong Wang (Beijing University of Posts and Telecommunications)
SegmentationConvolutional Neural NetworkSupervised Fine-TuningImageBiomedical DataComputed Tomography
🎯 What it does: In the semi-supervised medical image segmentation task, a Semantic Knowledge Complementary Decoupling Framework (SKCDF) is proposed, which trains the encoder, labeled decoder, and unlabeled decoder separately. It utilizes labeled data to guide pseudo-label generation and enriches labeled features with unlabeled data, while introducing an auxiliary balanced segmentation head to enhance the performance of minority classes.
A Simple Data Augmentation for Feature Distribution Skewed Federated Learning
Yunlu Yan (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
ClassificationSegmentationFederated LearningConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A random distribution normalization method based on input-level data augmentation (FedRDN) is proposed in federated learning, which alleviates feature distribution shift and enhances model generalization by randomly injecting global statistical information into local samples.
A Simple yet Effective Layout Token in Large Language Models for Document Understanding
Zhaoqing Zhu (Alibaba Group), Ji Zhang (Alibaba Group)
TransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: A scheme is proposed to compress document layout information into a single token and interleave it with text input to LLMs (LayTokenLLM), avoiding additional positional space waste by sharing position IDs.
A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs
Wangbo Zhao (National University of Singapore), Yang You (National University of Singapore)
Computational EfficiencyKnowledge DistillationTransformerVision Language ModelMultimodality
🎯 What it does: A training-independent method called SGL is proposed, which utilizes the full-layer attention information of a small-scale VLM to guide the visual token pruning of a large-scale VLM, and enhances inference efficiency through early exit of the small VLM when necessary.
A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets
David Mildenberger (Technical University of Munich), Martin J. Menten (Technical University of Munich)
ClassificationRepresentation LearningContrastive LearningImageBiomedical Data
🎯 What it does: This study investigates the performance of supervised contrastive learning on imbalanced binary classification datasets and proposes two improvement methods.
A Theory of Learning Unified Model via Knowledge Integration from Label Space Varying Domains
Dexuan Zhang (University of Tokyo), Tatsuya Harada (University of Tokyo)
Domain AdaptationAdversarial AttackImage
🎯 What it does: This paper proposes a unified model that addresses the multi-source semi-supervised open set domain adaptation problem using joint error theory and multi-class PU learning, and introduces attention feature generation to achieve source-free adaptation in the absence of source data.
A Unified Approach to Interpreting Self-supervised Pre-training Methods for 3D Point Clouds via Interactions
Qiang Li (Tongji University), Wen Shen (Tongji University)
ClassificationSegmentationRepresentation LearningGraph Neural NetworkPoint Cloud
🎯 What it does: This paper explores the common mechanisms of different self-supervised pre-training methods for 3D point cloud networks through game theory interaction analysis. Based on this, it proposes a loss term that directly enhances high-order interactions and suppresses low-order interactions, achieving performance comparable to traditional pre-training without the need for large-scale pre-training.
A Unified Framework for Heterogeneous Semi-supervised Learning
Marzi Heidari (Carleton University), Yuhong Guo (Carleton University)
ClassificationDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes the Heterogeneous Semi-Supervised Learning (HSSL) task and designs a unified 2C class fine-grained classification framework Uni-HSSL to simultaneously utilize labeled and unlabeled data from different domains for training;
A Unified Image-Dense Annotation Generation Model for Underwater Scenes
Hongkai Lin (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
SegmentationGenerationDepth EstimationTransformerSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a unified text-to-image and dense annotation generation model called TIDE, which can synthesize realistic underwater images along with their corresponding depth maps and semantic masks based solely on text prompts.
A Unified Latent Schrodinger Bridge Diffusion Model for Unsupervised Anomaly Detection and Localization
Shilhora Akshay (Indian Institute of Technology), Vineeth N Balasubramanian (Indian Institute of Technology)
Anomaly DetectionDiffusion modelAuto EncoderImageStochastic Differential Equation
🎯 What it does: A latent space diffusion model LASB based on linear Schrodinger Bridge is proposed for unsupervised anomaly detection and localization.
A Unified Model for Compressed Sensing MRI Across Undersampling Patterns
Armeet Singh Jatyani (California Institute of Technology), Anima Anandkumar (California Institute of Technology)
RestorationCompressionDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A unified neural operator model is proposed for compressed sensing MRI reconstruction, capable of maintaining performance across different sampling patterns and image resolutions.
A Unified, Resilient, and Explainable Adversarial Patch Detector
Vishesh Kumar (Indian Institute of Science Education and Research Bhopal), Akshay Agarwal (Indian Institute of Science Education and Research Bhopal)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: The research proposes a universal, robust, and interpretable adversarial patch detector called AdvPatchXAI.
A Universal Scale-Adaptive Deformable Transformer for Image Restoration across Diverse Artifacts
Xuyi He (South China University of Technology), Hui Ji (National University of Singapore)
RestorationTransformerImage
🎯 What it does: A scalable adaptive deformable Transformer (SADT) is proposed to eliminate structured artifacts such as rain, moiré patterns, and banding.
A3: Few-shot Prompt Learning of Unlearnable Examples with Cross-Modal Adversarial Feature Alignment
Xuan Wang (National University of Defense Technology), Cheng-zhong Xu (University of Macau)
ClassificationRecognitionData-Centric LearningMeta LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In the few-shot prompt learning scenario, the impact of unlearnable examples (UE) on the model is studied, and an adaptive UE generation framework and cross-modal adversarial feature alignment method A^3 are proposed to enhance the model's robustness in learning from UEs.
A4A: Adapter for Adapter Transfer via All-for-All Mapping for Cross-Architecture Models
Keyu Tu (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
GenerationData SynthesisTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes the A4A (Adapter for Adapter) framework, which enables seamless migration of attention-based adapters between different architectures (such as U-Net and Transformer) through all-for-all mapping.
AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP
Wenxin Ma (University of Science and Technology of China), S.Kevin Zhou (University of Science and Technology of China)
Anomaly DetectionTransformerContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposes the AA-CLIP two-stage adaptation method, transforming CLIP to achieve zero-shot anomaly detection and localization.
ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects
Woojin Lee (Korea Advanced Institute of Science and Technology), Munchurl Kim (Korea Advanced Institute of Science and Technology)
Object DetectionImage
🎯 What it does: A weakly supervised rotation object detection framework ABBSPO based on horizontal boxes (HBox) is proposed.
ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance
Yu-Cheng Chiu (National Chengchi University), Yan-Tsung Peng (National Chengchi University)
Image TranslationRestorationTransformerImage
🎯 What it does: An auxiliary dual-modal cross-domain Transformer named ABC-Former is proposed to improve the white balance correction of sRGB images.
AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers
Sherwin Bahmani (University of Toronto), Sergey Tulyakov (Snap Inc.)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: This paper proposes a 3D camera control method AC3D for video diffusion transformers, which can accurately generate videos with specified camera trajectories.
ACAttack: Adaptive Cross Attacking RGB-T Tracker via Multi-Modal Response Decoupling
Xinyu Xiang (Wuhan University), Jiayi Ma (Wuhan University)
Object TrackingAdversarial AttackVideoMultimodality
🎯 What it does: An adaptive cross-modal attack framework, ACAttack, is proposed for RGB-T multimodal trackers, capable of generating multimodal adversarial patches that can be deployed in both digital and physical domains, inducing tracker failure.
Acc3D: Accelerating Single Image to 3D Diffusion Models via Edge Consistency Guided Score Distillation
Kendong Liu, Junhui Hou
GenerationData SynthesisComputational EfficiencyKnowledge DistillationAdversarial AttackDiffusion modelScore-based ModelImagePoint CloudStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes Acc3D, a method for accelerating single-image to 3D diffusion models using edge consistency-guided distillation and decoupled adversarial regularization.
Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition
Zhiyuan Chen (Peking University), Yufei Ma (Peking University)
GenerationComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: This paper proposes a training-independent Incremental Calibration Cache method (ICC) to accelerate the inference process of the Diffusion Transformer (DiT) model.
Accelerating Multimodal Large Language Models by Searching Optimal Vision Token Reduction
Shiyu Zhao (Rutgers University), Licheng Yu (Meta)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This study investigates a prompt-aware visual token compression method that achieves inference acceleration by automatically searching for the optimal visual token reduction strategy within a multimodal large language model (MLLM);
Accurate Differential Operators for Hybrid Neural Fields
Aditya Chetan (Cornell University), Bharath Hariharan (Cornell University)
Supervised Fine-TuningNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper addresses the issue of high-frequency noise gradients generated during the automatic differentiation (AD) process in Hybrid Neural Fields, proposing a derivative operator based on local polynomial fitting and a self-supervised fine-tuning method to recover accurate gradients.
Accurate Scene Text Recognition with Efficient Model Scaling and Cloze Self-Distillation
Andrea Maracani (Samsung Research and Development Institute), Mete Ozay (Samsung Research and Development Institute)
RecognitionComputational EfficiencyKnowledge DistillationTransformerImage
🎯 What it does: This paper conducts a model scaling analysis for Scene Text Recognition (STR) and proposes the Cloze Self-Distillation (CSD) technique to suppress label noise in real datasets, introducing Differential Cross-Attention in the decoder for more efficient context modeling.
ACE: Anti-Editing Concept Erasure in Text-to-Image Models
Zihao Wang, Wangmeng Zuo
GenerationOptimizationConvolutional Neural NetworkImage
🎯 What it does: The paper proposes a new method to address a specific problem in computer vision, with details not elaborated.
ACL: Activating Capability of Linear Attention for Image Restoration
Yubin Gu (Xiamen University), Xiaoshuai Sun (National University of Singapore)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A new image restoration model called ACL is proposed, which integrates linear attention and the Mamba structure to build an efficient encoder-decoder network.
Acquire and then Adapt: Squeezing out Text-to-Image Model for Image Restoration
Junyuan Deng (Honor Device Co), Zhenyao Wu (Honor Device Co)
RestorationGenerationSuper ResolutionDiffusion modelImage
🎯 What it does: Using the pre-trained large-scale text-to-image model Flux, we constructed an unlabeled data generation pipeline called FluxGen and designed a lightweight adapter called FluxIR to control Flux for image restoration tasks.
Action Detail Matters: Refining Video Recognition with Local Action Queries
Mengmeng Wang (Zhejiang University of Technology), Yong Liu (Zhejiang University)
RecognitionTransformerVideo
🎯 What it does: A FocusVideo framework is proposed, which combines global video features with local action queries to enhance video action recognition accuracy.
Activating Sparse Part Concepts for 3D Class Incremental Learning
Zhenya Tian (University of Chinese Academy of Sciences), Haiyong Jiang (University of Chinese Academy of Sciences)
ClassificationObject DetectionTransformerPoint Cloud
🎯 What it does: This study focuses on 3D category incremental learning and proposes a framework called ILPC based on the concept of sparse activation components and task-level fusion.
Active Data Curation Effectively Distills Large-Scale Multimodal Models
Vishaal Udandarao (University of Tübingen), Olivier J. Henaff (Google DeepMind)
RetrievalKnowledge DistillationContrastive LearningMultimodality
🎯 What it does: A framework for efficient distillation of large multimodal models through Active Data Selection (ACID) is proposed, which is further combined with traditional Knowledge Distillation (KD) to obtain the ACED model, significantly improving the performance of small models in zero-shot classification and image-text retrieval.
Active Event-based Stereo Vision
Jianing Li (Tsinghua University), Xiangyang Ji (Tsinghua University)
Depth EstimationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an active event-based stereo vision system that combines a binocular event camera and an infrared projector to achieve high-speed depth perception.
Active Hyperspectral Imaging Using an Event Camera
Bohan Yu (Peking University), Imari Sato (National Institute of Informatics)
CompressionOptimizationImageVideo
🎯 What it does: Developed an active hyperspectral imaging system based on event cameras, achieving real-time capture and significantly reducing bandwidth.
ActiveGAMER: Active GAussian Mapping through Efficient Rendering
Liyan Chen (OPPO US Research Center), Yi Xu (OPPO US Research Center)
OptimizationRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingPoint CloudMesh
🎯 What it does: A real-time high-quality geometric and rendering reconstruction system called ActiveGAMER based on 3D Gaussian projection has been developed, capable of operating in unknown environments.
AdaCM^2: On Understanding Extremely Long-Term Video with Adaptive Cross-Modality Memory Reduction
Yuanbin Man (University of Texas at Arlington), Miao Yin (University of Texas at Arlington)
RecognitionCompressionTransformerLarge Language ModelVideoText
🎯 What it does: Proposes the AdaCM 2 framework, which utilizes autoregressive methods to perform cross-attention between video visual features and text prompts in Q-Former, and adaptively compresses visual memory based on this attention to achieve efficient understanding of extremely long videos.
AdaDARE-gamma: Balancing Stability and Plasticity in Multi-modal LLMs through Efficient Adaptation
Jingyi Xie (Information Research Center of Military Science), Wenpeng Hu (Beihang University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A framework for efficiently adapting multi-modal large language models (MLLMs) is proposed, aimed at alleviating catastrophic forgetting during the fine-tuning process.
AdaMMS: Model Merging for Heterogeneous Multimodal Large Language Models with Unsupervised Coefficient Optimization
Yiyang Du (Tsinghua University), Yang Liu (Tsinghua University)
OptimizationHyperparameter SearchTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: The AdaMMS method is proposed to achieve parameter merging for heterogeneous multimodal large language models.
AdaptCMVC: Robust Adaption to Incremental Views in Continual Multi-view Clustering
Jing Wang (Beijing Jiaotong University), Michael C. Kampffmeyer (UiT Arctic University of Norway)
Domain AdaptationAuto EncoderImage
🎯 What it does: Proposes the AdaptCMVC method, modeling the issues of view increment, noise interference, and forgetting in the context of Continuous Multi-View Clustering (CMVC).
Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning
Takuma Fukuda (Chiba University), Kazuhiko Kawamoto (Chiba University)
ClassificationComputational EfficiencyTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes a sample-free, class-incremental learning framework called ACMap, which achieves fixed inference time by merging task-specific adapters.
Adapting Dense Matching for Homography Estimation with Grid-based Acceleration
Kaining Zhang (Wuhan University), Paolo Favaro (University of Bern)
Image TranslationOptimizationComputational EfficiencyConvolutional Neural NetworkOptical FlowImage
🎯 What it does: GFNet is proposed, achieving high-resolution camera plane transformation estimation through sparse grid flow regression.
Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception
Baixuan Lv (Tsinghua University), Shu-Tao Xia (Tsinghua University)
RecognitionTransformerSupervised Fine-TuningVideoPoint Cloud
🎯 What it does: Transfer the pre-trained static 3D point cloud model to 4D point cloud videos, proposing Cross-frame Spatio-temporal Adaptation (CSA) to capture short-term and long-term spatio-temporal dynamics.
Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration
Chao Wang (University of Technology Sydney), Yi Yang (University of Technology Sydney)
RestorationTransformerVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes DiffRes, which achieves unified processing of various image restoration tasks by utilizing the feature differential instructions (FDI) generated by BLIP-2 and injecting them into a stable diffusion model with a lightweight adapter.
Adapting to Observation Length of Trajectory Prediction via Contrastive Learning
Ruiqi Qiu (Northeastern University), Yi Cen (Northeastern University)
Recurrent Neural NetworkContrastive LearningTime SeriesSequential
🎯 What it does: This paper proposes an adaptive method CLLS and a lightweight RNN model RNLS for trajectory prediction that aims to improve prediction accuracy under varying observation lengths.
Adapting to the Unknown: Training-Free Audio-Visual Event Perception with Dynamic Thresholds
Eitan Shaar (Bar-Ilan University), Lior Wolf (Tel-Aviv University)
RecognitionContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Proposed AV2A, a training-free, open-vocabulary audio-video event perception method that achieves cross-modal event localization through score-level fusion and dynamic thresholds.
Adaptive Dropout: Unleashing Dropout across Layers for Generalizable Image Super-Resolution
Hang Xu (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes Adaptive Dropout, a novel regularization method for blind image super-resolution (blind SR), which can adaptively incorporate Dropout in the intermediate layers of the network and enhance the model's generalization ability through a hierarchical annealing strategy.
Adaptive Keyframe Sampling for Long Video Understanding
Xi Tang (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
OptimizationTransformerVision Language ModelVideoBenchmark
🎯 What it does: An Adaptive Keyframe Sampling (AKS) algorithm is designed as a pluggable preprocessing module to select the most informative frames as the visual context for MLLM in long video understanding.
Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding
Han Xiao (CUHK MMLab), Hongsheng Li (CUHK MMLab)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: An adaptive generation pipeline based on multiple markup languages (Markdown, LaTeX, HTML, JSON, TikZ, Plain Text) has been designed to convert visual documents into structured code and serve as an intermediate reasoning basis, thereby enhancing the understanding and question-answering performance of visual documents.
Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training
Myunsoo Kim (Korea University), Byung-Jun Lee (Korea University)
GenerationOptimizationComputational EfficiencyDiffusion modelImage
🎯 What it does: An adaptive non-uniform time step sampling method is proposed to accelerate the training of diffusion models by dynamically adjusting the sampling probability of time steps.
Adaptive Parameter Selection for Tuning Vision-Language Models
Yi Zhang (Beihang University), Shi-Min Hu (Tsinghua University)
TransformerVision Language ModelImageMultimodality
🎯 What it does: This paper proposes CLIP-AST, which utilizes the second-order gradient adaptive learning rate of AdamW to automatically select the most important sub-layer parameters in the CLIP model for fine-tuning, without the need for manual localization or introducing additional parameters.
Adaptive Part Learning for Fine-Grained Generalized Category Discovery: A Plug-and-Play Enhancement
Qiyuan Dai (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
ClassificationRecognitionTransformerContrastive LearningImage
🎯 What it does: An adaptive part learning module (APL) for self-supervised unlabeled learning is proposed, which can be directly integrated into existing fine-grained category discovery frameworks to enhance performance.
Adaptive Rectangular Convolution for Remote Sensing Pansharpening
Xueyang Wang (University of Electronic Science and Technology of China), Liang-Jian Deng (University of Electronic Science and Technology of China)
RestorationConvolutional Neural NetworkImage
🎯 What it does: An Adaptive Rectangular Convolution (ARConv) module is proposed for pyramid resolution synthesis in remote sensing image fusion.
Adaptive Unimodal Regulation for Balanced Multimodal Information Acquisition
Chengxiang Huang (Beijing University of Posts and Telecommunications), Di Hu (Renmin University of China)
OptimizationVideoMultimodalityAudio
🎯 What it does: A method called Information Retrieval Regulation (InfoReg) is proposed to balance information retrieval in multimodal learning, particularly during the early learning phase (referred to as the primary learning window), by suppressing the information retrieval speed of information-rich modalities to promote information retrieval in information-scarce modalities.
ADD: Attribution-Driven Data Augmentation Framework for Boosting Image Super-Resolution
Ze-Yu Mi (Nanjing University), Yu-Bin Yang (Nanjing University)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: A framework for attention data augmentation based on attribution, ADD/ADD+, and novel Calibrated Attribution Maps (CAM) is proposed to improve super-resolution training in low-level vision tasks.
AdMiT: Adaptive Multi-Source Tuning in Dynamic Environments
Xiangyu Chang (University of California), Amit Roy-Chowdhury (University of California)
ClassificationSegmentationDomain AdaptationTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes an efficient tuning (PET) module that selects and integrates pre-trained parameters to achieve adaptability of the Transformer model in dynamic environments.
ADU: Adaptive Detection of Unknown Categories in Black-Box Domain Adaptation
Yushan Lai (Sun Yat-Sen University), Zhiyu Ye (Sun Yat-Sen University)
Domain AdaptationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: The ADU framework is proposed for black-box domain adaptation in situations where the target domain has unknown classes, enabling knowledge transfer and unknown class detection solely based on the outputs of the source domain black-box predictor.
Adv-CPG: A Customized Portrait Generation Framework with Facial Adversarial Attacks
Junying Wang (Northwestern Polytechnical University), Yuan Yuan (Northwestern Polytechnical University)
GenerationSafty and PrivacyAdversarial AttackDiffusion modelImageMultimodality
🎯 What it does: Proposed Adv-CPG, a framework that achieves facial privacy protection while generating personalized portraits;
Advancing Adversarial Robustness in GNeRFs: The IL2-NeRF Attack
Nicole Meng (Tufts University), Yingjie Lao (Tufts University)
GenerationDepth EstimationAdversarial AttackNeural Radiance FieldImage
🎯 What it does: An iterative adversarial attack method based on L2 norm, IL2-NeRF, is proposed for General Neural Radiance Fields (GNeRF), which can apply overall smooth perturbations on source view images, thereby disrupting the generated 3D views and depth.
Advancing Generalizable Tumor Segmentation with Anomaly-Aware Open-Vocabulary Attention Maps and Frozen Foundation Diffusion Models
Yankai Jiang (Shanghai AI Laboratory), Xiaosong Wang (Shanghai AI Laboratory)
SegmentationAnomaly DetectionDiffusion modelContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The DiffuGTS framework is proposed, which generates anomaly-aware open attention maps using the internal representations of a frozen medical diffusion model, and achieves zero-shot tumor segmentation through latent space pseudo-healthy reconstruction and pixel/feature residual learning.