ICCV 2025 Papers — Page 3
IEEE/CVF International Conference on Computer Vision · 2701 papers
AVAM: a Universal Training-free Adaptive Visual Anchoring Embedded into Multimodal Large Language Model for Multi-image Question Answering
Kang Zeng (Hunan University), Zhiyong Li (Hunan University)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: An Adaptive Visual Anchoring method is proposed, which compresses visual tokens in multi-image visual question answering (MVQA) without training, enhancing the reasoning accuracy and efficiency of multimodal large language models (MLLMs).
Avat3r: Large Animatable Gaussian Reconstruction Model for High-fidelity 3D Head Avatars
Tobias Kirschstein (Technical University of Munich), Shunsuke Saito (Meta Reality Labs)
GenerationTransformerGenerative Adversarial NetworkImageVideo
🎯 What it does: A method called Avat3r is proposed, which can generate high-fidelity and animatable 3D avatars using only four head images, simplifying the capture process.
AVTrustBench: Assessing and Enhancing Reliability and Robustness in Audio-Visual LLMs
Sanjoy Chowdhury (University of Maryland), Dinesh Manocha (University of Maryland)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodalityBenchmarkAudio
🎯 What it does: A new evaluation benchmark for audio-video large language models (AVLLM) called AVTRUSTBENCH is proposed, and a systematic evaluation of 16 state-of-the-art AVLLMs is conducted; at the same time, a model-agnostic and robust preference optimization method called CAVPref is introduced.
Axis-level Symmetry Detection with Group-Equivariant Representation
Wongyun Yu (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a framework for detecting axial symmetry using the dihedral group (D_N) equivariant representation, which directly predicts the geometric parameters of reflection axes and rotation axes.
B-VLLM: A Vision Large Language Model with Balanced Spatio-Temporal Tokens
Zhuqiang Lu (University of Sydney), Kun Hu (Edith Cowan University)
TransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: Proposes the B-VLLM framework, which utilizes text-conditioned adaptive frame selection, temporal frame merging, and spatial visual token sampling/merging to achieve dynamic balance of video spatiotemporal information and control the number of visual tokens.
BabyVLM: Data-Efficient Pretraining of VLMs Inspired by Infant Learning
Shengao Wang (Boston University), Boqing Gong (Boston University)
RecognitionGenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Design the BabyVLM framework, utilizing infant learning patterns for data-efficient visual-language model pre-training, and propose a multi-task evaluation benchmark aimed at the infant stage.
Back on Track: Bundle Adjustment for Dynamic Scene Reconstruction
Weirong Chen (Technical University of Munich), Daniel Cremers (Technical University of Munich)
Depth EstimationOptimizationTransformerSimultaneous Localization and MappingVideo
🎯 What it does: Utilizing a 3D point tracker to separate camera motion from object motion, then performing bundle adjustment on all points and achieving global depth refinement through a scale grid, enabling dense reconstruction of dynamic scenes in arbitrary videos.
Backdoor Attacks on Neural Networks via One-Bit Flip
Xiang Li (George Mason University), Qiang Zeng (George Mason University)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes SOLEFLIP, a backdoor injection method that achieves this by flipping a single bit in quantized deep neural networks.
Backdoor Defense via Enhanced Splitting and Trap Isolation
Hongrui Yu (Beihang University), Chengbin Sun (Beihang University)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A backdoor defense framework named ESTI is proposed, which utilizes data tampered by attackers along with a 'poisoning model' to partition the dataset. It then isolates the identified backdoor samples into a dedicated class using 'trap labels', ultimately training a model that is ineffective against backdoors while maintaining high accuracy.
Backdoor Mitigation by Distance-Driven Detoxification
Shaokui Wei (Chinese University of Hong Kong), Hongyuan Zha (Chinese University of Hong Kong)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a post-training debiasing method called D3, which aims to remove backdoors in pre-trained models by maximizing the distance between model weights and the original backdoored model while maintaining low loss on clean data.
Backdooring Self-Supervised Contrastive Learning by Noisy Alignment
Tuo Chen (Southeast University), Jian Liu (Ant Group)
Representation LearningAdversarial AttackContrastive LearningImage
🎯 What it does: This paper proposes a data poisoning backdoor attack based on Noisy Alignment, which aligns noisy images with target reference images using randomly cropped reverse layouts, thereby implanting a powerful and covert backdoor in unsupervised contrastive learning.
Background Invariance Testing According to Semantic Proximity
Zukang Liao (University of Oxford), Min Chen (University of Oxford)
ClassificationRecognitionImage
🎯 What it does: A visualization-based framework for background invariance testing is proposed, using associated ontologies to extend keywords for background scene sampling to evaluate the background invariance of machine learning models.
BadVideo: Stealthy Backdoor Attack against Text-to-Video Generation
Ruotong Wang (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
GenerationData SynthesisAdversarial AttackLarge Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Proposed and implemented the BadVideo framework, utilizing the spatiotemporal redundancy information from text-to-video generation models for covert backdoor attacks;
Baking Gaussian Splatting into Diffusion Denoiser for Fast and Scalable Single-stage Image-to-3D Generation and Reconstruction
Yuanhao Cai (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a single-stage 3D Gaussian Splatting diffusion model called DiffusionGS, which can generate 3D objects and scenes from a single view without the need for depth estimation.
Balanced Image Stylization with Style Matching Score
Yuxin Jiang (National University of Singapore), Mike Zheng Shou
Image TranslationGenerationOptimizationDiffusion modelImage
🎯 What it does: Proposes the Style Matching Score (SMS) optimization method, achieving image style transfer based on diffusion models while balancing style alignment and content preservation.
Balanced Sharpness-Aware Minimization for Imbalanced Regression
Yahao Liu (University of Electronic Science and Technology of China), Wen Li (University of Electronic Science and Technology of China)
Depth EstimationOptimizationConvolutional Neural NetworkImage
🎯 What it does: Proposes Balanced Sharpness-Aware Minimization (BSAM) to address the generalization degradation problem caused by data imbalance in regression tasks.
Balancing Conservatism and Aggressiveness: Prototype-Affinity Hybrid Network for Few-Shot Segmentation
Tianyu Zou (Wuhan University of Technology), Yi Rong (Wuhan University of Technology)
RecognitionSegmentationTransformerImage
🎯 What it does: This study focuses on few-shot semantic segmentation and proposes the Prototype-Affinity Hybrid Network (PAHNet), which improves foreground recognition and background suppression by combining conservative prototype learning with aggressive affinity learning, utilizing the PFE and ASC modules.
Balancing Task-invariant Interaction and Task-specific Adaptation for Unified Image Fusion
Xingyu Hu (Harbin Institute of Technology), Jiayi Ma (Wuhan University)
RestorationTransformerImageMultimodalityMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A unified image fusion framework TITA is proposed, which balances task-invariant interaction and task-specific adaptation, enabling the completion of various fusion tasks without the need for task identification.
BANet: Bilateral Aggregation Network for Mobile Stereo Matching
Gangwei Xu (Huazhong University of Science and Technology), Xin Yang (Optics Valley Laboratory)
Depth EstimationAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A Bidirectional Aggregation Network (BANet) is proposed, which achieves high-quality stereo matching on mobile devices using only 2D convolutions, capable of preserving detail and edge information.
BASIC: Boosting Visual Alignment with Intrinsic Refined Embeddings in Multimodal Large Language Models
Jianting Tang (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: For multimodal large language models (MLLM), a BASIC method is proposed, which utilizes the refined visual embeddings from the earlier layers of the LLM as direct visual supervision to guide the visual projector in generating higher quality initial visual embeddings, thereby enhancing the alignment between vision and language.
BATCLIP: Bimodal Online Test-Time Adaptation for CLIP
Sarthak Maharana, Yunhui Guo (MIT-IBM Watson AI Lab)
Domain AdaptationTransformerContrastive LearningImageText
🎯 What it does: A dual-modal online testing adaptation method called BATCLIP is proposed to enhance the robustness of CLIP under image distortion.
Bayesian-Inspired Space-Time Superpixels
Kent Gauen (Purdue University), Stanley Chan (Purdue University)
Object TrackingSegmentationOptical FlowVideo
🎯 What it does: This paper proposes a Bayesian heuristic spatiotemporal superpixel segmentation method named BIST, which achieves real-time performance while maintaining high segmentation quality.
Benchmarking and Learning Multi-Dimensional Quality Evaluator for Text-to-3D Generation
Yujie Zhang (Shanghai Jiao Tong University), Yiling Xu (Shanghai Jiao Tong University)
GenerationData SynthesisPrompt EngineeringTextMeshBenchmark
🎯 What it does: This paper constructs a multi-dimensional text-to-3D generation evaluation benchmark called MATE-3D and proposes a multi-dimensional quality evaluator, HyperScore, based on conditional feature fusion and hypernetworks.
Benchmarking Burst Super-Resolution for Polarization Images: Noise Dataset and Analysis
Inseung Hwang (Korea Advanced Institute of Science and Technology), Min H. Kim (Korea Advanced Institute of Science and Technology)
RestorationSuper ResolutionImageBenchmark
🎯 What it does: This paper presents two publicly available datasets, PolarNS and PolarBurstSR, specifically designed for noise statistics and super-resolution of polarized images, and conducts theoretical modeling and empirical validation of polarized noise.
Benchmarking Egocentric Visual-Inertial SLAM at City Scale
Anusha Krishnan (ETH Zurich), Marc Pollefeys (ETH Zurich)
Autonomous DrivingOptimizationRobotic IntelligenceSimultaneous Localization and MappingVideoMultimodalityBenchmark
🎯 What it does: A new visual-inertial SLAM benchmark (LaMAria) for urban scale and head-mounted multi-sensor data is proposed, along with an evaluation method.
Benchmarking Multimodal CoT Reward Model Stepwise by Visual Program
Minghe Gao (Zhejiang University), Yueting Zhuang (Zhejiang University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelImageVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Developed SVIP, a framework for automatically generating and annotating multi-dimensional chain-of-thought (CoT) steps based on visual programming, and trained a step-level reward model SVIP-Reward to enhance the reasoning and answer filtering of multi-modal large language models (MLLMs).
Benchmarking Multimodal Large Language Models Against Image Corruptions
Xinkuan Qiu (Institute of Information Engineering, Chinese Academy of Sciences), Shiguang Shan (Institute of Computing Technology, Chinese Academy of Sciences)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A robustness evaluation benchmark MLLM-IC for multimodal large language models under image distortion conditions is proposed.
Benefit From Seen: Enhancing Open-Vocabulary Object Detection by Bridging Visual and Textual Co-Occurrence Knowledge
Yanqi Li (Beihang University), Tao Ren (Institute of Software Chinese Academy of Sciences)
Object DetectionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A plugin framework named CODet is proposed, which first extracts visually co-occurring objects from images, then uses a large language model (LLM) to validate the co-occurrence relationships, and finally injects the obtained co-occurrence pseudo-labels into a VLM-based open vocabulary object detection model to enhance recognition performance for unknown categories.
Beyond [cls]: Exploring the True Potential of Masked Image Modeling Representations
Marcin Przewięźlikowski (Jagiellonian University), Bartosz Zieliński (Jagiellonian University)
Representation LearningTransformerAuto EncoderImage
🎯 What it does: A deep analysis of the attention distribution in Masked Image Modeling (MIM) models is conducted, and a lightweight Selective Aggregation method is proposed, significantly improving MIM's performance on various downstream tasks without fine-tuning.
Beyond Blur: A Fluid Perspective on Generative Diffusion Models
Grzegorz Gruszczynski (Samsung AI Center), Przemyslaw Musialski (New Jersey Institute of Technology)
RestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: A physical-driven image erosion process based on the convection-diffusion equation is proposed for forward perturbation of diffusion models.
Beyond Brain Decoding: Visual-Semantic Reconstructions to Mental Creation Extension Based on fMRI
Haodong Jing (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
GenerationRetrievalLarge Language ModelMixture of ExpertsDiffusion modelImageTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: Proposes NeuroCreat, a brain multimodal architecture that decodes fMRI signals into visual-semantic representations and can generate creative content.
Beyond Isolated Words: Diffusion Brush for Handwritten Text-Line Generation
Gang Dai (South China University of Technology), Shuicheng Yan (National University of Singapore)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelImageText
🎯 What it does: This paper proposes DiffBrush, a diffusion model for generating handwritten text lines; it achieves style imitation of the overall text line and control over character readability through content-decoupled style learning and a multi-scale content discriminator.
Beyond Label Semantics: Language-Guided Action Anatomy for Few-shot Action Recognition
Zefeng Qian (Shanghai Jiao Tong University), Hong Sun (E-surfing Vision Technology Co., Ltd)
RecognitionTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes the LGA framework, which utilizes LLM to deconstruct action labels and divides videos into atomic phases for few-shot action recognition.
Beyond Losses Reweighting: Empowering Multi-Task Learning via the Generalization Perspective
Hoang Phan (New York University), Trung Le (Monash University)
Object DetectionSegmentationAutonomous DrivingSupervised Fine-TuningImage
🎯 What it does: A multi-task learning framework based on weight perturbation is proposed, which alleviates task conflicts and enhances model generalization by adjusting the gradient norm.
Beyond Low-Rank Tuning: Model Prior-Guided Rank Allocation for Effective Transfer in Low-Data and Large-Gap Regimes.
Chuyan Zhang (Shanghai Jiao Tong University), Yun Gu (Shanghai Jiao Tong University)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A hierarchical low-rank adaptive method SR-LoRA based on the stable rank of pre-trained model weights is proposed, combined with stochastic local updates for lightweight implementation.
Beyond Next-Token: Next-X Prediction for Autoregressive Visual Generation
Sucheng Ren, Liang-Chieh Chen
GenerationTransformerFlow-based ModelAuto EncoderImage
🎯 What it does: The xAR framework is proposed, extending autoregressive generation from traditional 'next token' prediction to 'next X' prediction, where X can be image patches, cells, sub-samples, scales, or entire images; and the model is trained using Noisy Context Learning to combat exposure bias.
Beyond One Shot, Beyond One Perspective: Cross-View and Long-Horizon Distillation for Better LiDAR Representations
Xiang Xu (Nanjing University of Aeronautics and Astronautics), Qingshan Liu (Nanjing University of Posts and Telecommunications)
Object DetectionSegmentationAutonomous DrivingKnowledge DistillationRepresentation LearningContrastive LearningImagePoint Cloud
🎯 What it does: The LiMA framework is proposed, which achieves more robust representation learning for LiDAR through cross-view aggregation, long-term feature propagation, and cross-sequence memory alignment for knowledge distillation from image to LiDAR.
Beyond Perspective: Neural 360-Degree Video Compression
Andy Regensky (Friedrich-Alexander-Universität Erlangen-Nürnberg), Andre Kaup (Friedrich-Alexander-Universität Erlangen-Nürnberg)
CompressionOptical FlowVideo
🎯 What it does: A neural network compression framework for 360-degree video has been developed, combining large-scale datasets, spherical reprojection enhancement, and positional encoding to achieve bidirectional optimization for both 360-degree and traditional perspective videos.
Beyond Pixel Uncertainty: Bounding the OoD Objects in Road Scenes
Huachao Zhu (Wuhan University), Yongchao Xu (Wuhan University)
Object DetectionAnomaly DetectionAutonomous DrivingImageVideo
🎯 What it does: A detection-suppression-based road anomaly detection framework called DetSeg is proposed, which suppresses ID boxes after generating all object boxes using an open object detector, retaining only OoD boxes for anomaly detection.
Beyond RGB: Adaptive Parallel Processing for RAW Object Detection
Shani Gamrian (Sony Research), Daisuke Iso (Sony Group Corporation)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Raw Adaptation Module (RAM), a parallel ISP preprocessing module for end-to-end training on RAW images to enhance object detection performance.
Beyond Simple Edits: Composed Video Retrieval with Dense Modifications
Omkar Thawakar (Mohamed bin Zayed University of AI), Salman Khan (Linköping University)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a fine-grained video retrieval task, constructs the Dense-WebVid-CoVR dataset, and designs a unified fusion model to achieve joint encoding of video, text descriptions, and dense modified text.
Beyond Single Images: Retrieval Self-Augmented Unsupervised Camouflaged Object Detection
Ji Du (Nankai University), Ping Li (Hong Kong Polytechnic University)
Object DetectionRetrievalTransformerContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised hidden object detection framework called RISE, which generates pseudo-labels using a retrieval-incremental strategy to train the COD model.
Beyond Spatial Frequency: Pixel-wise Temporal Frequency-based Deepfake Video Detection
Taehoon Kim (Chung Ang University), Jongwon Choi (Chung Ang University)
Anomaly DetectionTransformerVideo
🎯 What it does: This paper proposes a deepfake video detection method that utilizes pixel-level temporal frequency features, capable of accurately capturing fine-grained temporal inconsistencies between video frames.
Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning in VLMs
Qizhe Zhang (Peking University), Shanghang Zhang (Peking University)
OptimizationComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: A training-free visual token pruning method based on visual encoder attention and similarity, called VisPruner, is proposed.
Beyond the Destination: A Novel Benchmark for Exploration-Aware Embodied Question Answering
Kaixuan Jiang (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
TransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A large-scale EQA benchmark dataset called EXPRESS-Bench has been constructed, and a Fine-EQA hybrid exploration model has been proposed.
Beyond the Frame: Generating 360deg Panoramic Videos from Perspective Videos
Rundong Luo (Cornell University), Wei-Chiu Ma (Cornell University)
GenerationData SynthesisDiffusion modelSimultaneous Localization and MappingVideo
🎯 What it does: A model called Argus is proposed for generating panoramic 360° videos from perspective videos.
Beyond the Limits: Overcoming Negative Correlation of Activation-Based Training-Free NAS
Haidong Kang (Northeastern University), Min Huang (Northeastern University)
Neural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new training-free neural architecture search method that improves upon the negative correlation issue in activation-based proxies.
Beyond Training: Dynamic Token Merging for Zero-Shot Video Understanding
Yiming Zhang (University of Science and Technology of China), Yining Sun (HFIPS)
RecognitionCompressionComputational EfficiencyTransformerVision Language ModelVideo
🎯 What it does: A zero-shot video understanding framework called DYTO is proposed, which can efficiently and semantically encode videos without additional fine-tuning.
Beyond Walking: A Large-Scale Image-Text Benchmark for Text-based Person Anomaly Search
Shuyu Yang (Xi'an Jiaotong University), Zhedong Zheng (University of Macau)
RetrievalAnomaly DetectionContrastive LearningImageVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes a text-driven pedestrian anomaly search task, constructs a large-scale Pedestrian Anomaly Behavior (PAB) visual-language benchmark, and introduces a Cross-Modal Pose-aware (CMP) framework.
BezierGS: Dynamic Urban Scene Reconstruction with Bezier Curve Gaussian Splatting
Zipei Ma (Fudan University), Li Zhang (Fudan University)
Autonomous DrivingGaussian SplattingVideoBenchmark
🎯 What it does: This paper achieves high-quality reconstruction and novel view rendering of large-scale urban dynamic scenes through a learnable Bézier curve Gaussian splatting method.
Bi-Level Optimization for Self-Supervised AI-Generated Face Detection
Mian Zou (Jiangxi University of Finance and Economics), Kede Ma (City University of Hong Kong)
Object DetectionOptimizationContrastive LearningImage
🎯 What it does: A self-supervised AI-generated face detection method based on dual-layer optimization is proposed, which trains a visual encoder using EXIF tags and facial manipulation tasks without any manually synthesized samples.
Bias in Gender Bias Benchmarks: How Spurious Features Distort Evaluation
Yusuke Hirota (NVIDIA), Chao-Han Huck Yang (NVIDIA)
Vision Language ModelImageBenchmark
🎯 What it does: This study investigates the interference of non-gender features (such as color, objects, and backgrounds) in the assessment of gender bias in visual language models (VLMs) and measures their impact on bias metrics through systematic perturbations of these features, proposing to report feature sensitivity alongside bias reporting.
Bias-Resilient Weakly Supervised Semantic Segmentation Using Normalizing Flows
Xianglin Qiu (XJTLU), Jimin Xiao (XJTLU)
SegmentationConvolutional Neural NetworkFlow-based ModelContrastive LearningImage
🎯 What it does: Using normalizing flow to model the pixel feature distribution of the entire dataset, constructing distribution-based CAM and reliable feature sampling, thereby enhancing the robustness and accuracy of weakly supervised semantic segmentation.
Bidirectional Likelihood Estimation with Multi-Modal Large Language Models for Text-Video Retrieval
Dohwan Ko (Korea University), Hyunwoo J. Kim (KAIST)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodality
🎯 What it does: This study proposes a bidirectional likelihood estimation retrieval framework (BLiM) based on a multimodal large language model (MLLM), aimed at addressing the retrieval misjudgment problem caused by candidate prior bias.
Bilateral Collaboration with Large Vision-Language Models for Open Vocabulary Human-Object Interaction Detection
Yupeng Hu (South China University of Technology), Xiangmin Xu (South China University of Technology)
Object DetectionTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: A bidirectional collaborative framework BC-HOI is proposed to achieve open-vocabulary human-object interaction (HOI) detection.
BillBoard Splatting (BBSplat): Learnable Textured Primitives for Novel View Synthesis
David Svitov (Universita degli Studi di Genova), Alessio Del Bue (Istituto Italiano di Tecnologia)
GenerationData SynthesisCompressionGaussian SplattingImage
🎯 What it does: This paper proposes a BBSplat method based on learnable texture plane primitives for achieving high-quality view synthesis with a low number of primitives.
Bitrate-Controlled Diffusion for Disentangling Motion and Content in Video
Xiao Li (Microsoft Research), Yan Lu (Microsoft Research)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: A self-supervised diffusion framework BCD based on low bitrate vector quantization is proposed to decompose videos into motion features and content features, and to achieve high-quality motion transfer and video generation using these separated features.
Blended Point Cloud Diffusion for Localized Text-guided Shape Editing
Etai Sella (Tel Aviv University), Hadar Averbuch-Elor (Cornell University)
RestorationSegmentationGenerationLarge Language ModelDiffusion modelTextPoint Cloud
🎯 What it does: Achieve fine-grained editing of local geometry in point clouds through text prompts, utilizing diffusion model-based local inpainting to modify specified areas while keeping the structure of the remaining parts unchanged.
Blind Noisy Image Deblurring Using Residual Guidance Strategy
Heyan Liu (Northeast Normal University), Tieyong Zeng (Beijing Normal-Hong Kong Baptist University)
RestorationImage
🎯 What it does: This paper addresses the problem of blind image deblurring with high noise and proposes a Residual Guidance Strategy (RGS) that utilizes coarse-scale information from a multi-scale image pyramid to guide the estimation of the blur kernel at finer scales, completing the joint optimization of denoising and deblurring within an unsupervised traditional framework.
Blind Video Super-Resolution based on Implicit Kernels
Qiang Zhu (University of Electronic Science and Technology of China), Bing Zeng (University of Electronic Science and Technology of China)
RestorationSuper ResolutionTransformerOptical FlowVideo
🎯 What it does: This paper studies blind video super-resolution (BVSR) and proposes the BVSR-IK model based on implicit kernels, which generates a multi-scale kernel dictionary using implicit neural representations and predicts coefficients through a recursive Transformer, achieving spatial correction and spatiotemporal alignment.
Blind2Sound: Self-Supervised Image Denoising without Residual Noise
Jiazheng Liu (Chinese Academy of Sciences), Hua Han (Chinese Academy of Sciences)
RestorationImage
🎯 What it does: We propose Blind2Sound, a self-supervised blind denoising framework capable of removing Poisson-Gaussian noise without residual noise.
BlinkTrack: Feature Tracking over 80 FPS via Events and Images
Yichen Shen (Zhejiang University), Guofeng Zhang (Zhejiang University)
Object TrackingConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowMultimodality
🎯 What it does: Proposes the BlinkTrack framework, which integrates event cameras and grayscale images for high-frequency point tracking.
BlueNeg: A 35mm Negative Film Dataset for Restoring Channel-Heterogeneous Deterioration
Hanyuan Liu (City University of Hong Kong), Tien-Tsin Wong (Monash University)
RestorationSupervised Fine-TuningImage
🎯 What it does: Designed and released the BlueNeg 35mm negative dataset for the study and evaluation of image restoration methods under channel heterogeneous degradation.
BokehDiff: Neural Lens Blur with One-Step Diffusion
Chengxuan Zhu (Peking University), Boxin Shi (Peking University)
RestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: BokehDiff proposes a diffusion model that requires only one-step inference, utilizing a physics-inspired self-attention module to achieve direct rendering from fully focused images to realistic bokeh effects.
Bokehlicious: Photorealistic Bokeh Rendering with Controllable Apertures
Tim Seizinger (Computer Vision Lab, CAIDAS, University of Wurzburg), Radu Timofte (Computer Vision Lab, CAIDAS, University of Wurzburg)
RestorationGenerationTransformerImageBenchmark
🎯 What it does: The Bokehlicious framework and the RealBokeh dataset are proposed for achieving controllable and realistic bokeh rendering.
Bolt3D: Generating 3D Scenes in Seconds
Stanislaw Szymanowicz (Google), Philipp Henzler (Google)
GenerationData SynthesisDiffusion modelPoint Cloud
🎯 What it does: Bolt3D proposes an instant 3D scene generation framework based on latent diffusion models, capable of directly generating interactive rendered 3D Gaussian scenes from one or more input images in less than seven seconds.
Boost 3D Reconstruction using Diffusion-based Monocular Camera Calibration
Junyuan Deng (Hong Kong University of Science and Technology), Ping Tan (Hong Kong University of Science and Technology)
GenerationPose EstimationDepth EstimationAutonomous DrivingDiffusion modelImage
🎯 What it does: A monocular camera calibration method based on diffusion models, DM-Calib, is proposed, which can estimate camera intrinsic parameters using only a single input image and apply the results to various 3D vision tasks.
Boosting Adversarial Transferability via Negative Hessian Trace Regularization
Yunfei Long (Harbin Engineering University), Huosheng Xu (Harbin Engineering University)
Adversarial AttackImage
🎯 What it does: A method for adversarial attacks using negative Hessian matrix trace regularization (NHTR) is proposed, which generates flatter loss peaks by suppressing curvature in all directions, thereby enhancing the transferability of adversarial samples.
Boosting Adversarial Transferability via Residual Perturbation Attack
Jinjia Peng (Hebei University), Yang Wang (Hefei University of Technology)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A residual perturbation attack (ResPA) is proposed, which utilizes the residual gradient to guide the generation of adversarial samples, enhancing transferability on black-box models.
Boosting Class Representation via Semantically Related Instances for Robust Long-Tailed Learning with Noisy Labels
Yuhang Li (Southeast University), Yuheng Jia (Southeast University)
ClassificationRepresentation LearningMixture of ExpertsContrastive LearningImage
🎯 What it does: A long-tail noisy label learning method based on instance similarity soft labels and shot-specific expert ensemble (IBC) is proposed, which dynamically adjusts label probabilities during training and targets different shot-specific experts.
Boosting Domain Generalized and Adaptive Detection with Diffusion Models: Fitness, Generalization, and Transferability
Boyong He (Xiamen University), Liaoni Wu (Xiamen University)
Object DetectionDomain AdaptationKnowledge DistillationDiffusion modelImage
🎯 What it does: A target detection framework based on diffusion models is proposed, combining single-step feature extraction, object center auxiliary branches, and consistency loss to achieve domain generalization and adaptive detection.
Boosting Generative Adversarial Transferability with Self-supervised Vision Transformer Features
Shangbo Wu (Beijing Institute of Technology), Yuanzhang Li (Beijing Institute of Technology)
GenerationAdversarial AttackTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A generative adversarial attack method named dSVA is proposed, which utilizes the dual features of self-supervised Vision Transformer (ViT) (global structural features from contrastive learning DINO and local texture features from masked image modeling MAE) to generate adversarial samples with high cross-model transferability.
Boosting MLLM Reasoning with Text-Debiased Hint-GRPO
Qihan Huang (Zhejiang University), Jie Song (Zhejiang University)
Large Language ModelReinforcement LearningPrompt EngineeringMultimodality
🎯 What it does: This paper improves the low data utilization and text bias issues of GRPO by introducing Hint-GRPO and text bias correction methods in MLLM inference.
Boosting Multi-View Indoor 3D Object Detection via Adaptive 3D Volume Construction
Runmin Zhang (Zhejiang University), Hui-Liang Shen (Zhejiang University)
Object DetectionConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes SGCDet, a multi-view indoor 3D object detection framework based on adaptive 3D volume construction, which can be trained using only 3D bounding boxes, avoiding reliance on complete geometric information.
Boosting Multimodal Learning via Disentangled Gradient Learning
Shicai Wei (University of Electronic Science and Technology of China), Yang Luo (University of Electronic Science and Technology of China)
ClassificationRecognitionSegmentationVideoMultimodalityAudio
🎯 What it does: A separation gradient learning framework (DGL) is proposed, which eliminates the optimization conflict between the encoder and the fusion module by truncating the cross-modal loss on the encoder's gradient and replacing it with a single-modal gradient.
Boosting Vision Semantic Density with Anatomy Normality Modeling for Medical Vision-language Pre-training
Weiwei Cao (Zhejiang University), Ling Zhang (Alibaba Group)
ClassificationSegmentationGenerationConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityComputed Tomography
🎯 What it does: The ViSD-Boost method is proposed to enhance the visual semantic density of medical visual-language pre-training through disease-level contrastive learning and anatomical normality modeling, addressing the alignment bias caused by the difference in semantic density between vision and text.
Bootstrap3D: Improving Multi-view Diffusion Model with Synthetic Data
Zeyi Sun (Shanghai Jiaotong University), Jiaqi Wang (Shanghai Jiaotong University)
GenerationData SynthesisLarge Language ModelDiffusion modelImageVideo
🎯 What it does: Proposes the Bootstrap3D framework, which utilizes video diffusion models and 3D perception LLM to automatically generate high-quality multi-view images and dense descriptions, training a multi-view diffusion model.
Bootstrapping Grounded Chain-of-Thought in Multimodal LLMs for Data-Efficient Model Adaptation
Jiaer Xia (Hong Kong Baptist University), Kaiyang Zhou (Hong Kong Baptist University)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelVision Language ModelImageMultimodalityTabularChain-of-Thought
🎯 What it does: This study explores how to efficiently adapt multimodal large language models (MLLM) to specialized visual tasks with limited data (such as charts, tables, receipts, etc.) and proposes a self-verification based Grounded Chain-of-Thought (GCoT) method.
Borrowing Eyes for the Blind Spot: Overcoming Data Scarcity in Malicious Video Detection via Cross-Domain Retrieval Augmentation
Rongpei Hong (University of Electronic Science and Technology of China), Fan Zhou (University of Electronic Science and Technology of China)
RetrievalAnomaly DetectionContrastive LearningImageVideoText
🎯 What it does: A cross-domain retrieval enhancement framework CRAVE is proposed, which uses image-text data to enhance malicious video detection and addresses the issue of scarce video data.
Boundary Probing for Input Privacy Protection When Using LMM Services
Xiaofei Hui (Lancaster University), Jun Liu (Lancaster University)
OptimizationSafty and PrivacyConvolutional Neural NetworkLarge Language ModelVideo
🎯 What it does: A method for protecting input visual privacy under black-box large model services is proposed, constructing the PABP framework and designing two strategies, GEP and PGP, to train an anonymization network that maintains LMM performance while minimizing privacy leakage.
BoxDreamer: Dreaming Box Corners for Generalizable Object Pose Estimation
Yuanhong Yu (Zhejiang University), Sida Peng (Zhejiang University)
Pose EstimationTransformerImage
🎯 What it does: Proposes the BoxDreamer framework for RGB sparse perspective estimation of arbitrary object 6D poses.
Breaking Grid Constraints: Dynamic Graph Reconstruction Network for Multi-organ Segmentation
Junhao Xiao (Chongqing University of Posts and Telecommunications), Bin Xiao (Jinan Inspur Data Technology Co., Ltd.)
SegmentationGraph Neural NetworkBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This study investigates the multi-organ segmentation problem and proposes a segmentation network called DGRNet based on dynamic graph reconstruction.
Breaking Rectangular Shackles: Cross-View Object Segmentation for Fine-Grained Object Geo-Localization
Qingwang Zhang (Shenzhen University), Yingying Zhu (Shenzhen University)
Object DetectionSegmentationTransformerImage
🎯 What it does: A cross-view object segmentation (CVOS) scheme is proposed, utilizing a Transformer encoder and SAM to achieve pixel-level fine geographic localization of query objects in satellite images.
Breaking the Encoder Barrier for Seamless Video-Language Understanding
Handong Li (Chinese Academy of Sciences), Jing Liu (Chinese Academy of Sciences)
TransformerVision Language ModelVideoTextMultimodality
🎯 What it does: We propose ELVA, an Encoder-free video large language model that completely relies on learning video-text interactions directly from raw video pixels without a visual encoder.
BridgeDepth: Bridging Monocular and Stereo Reasoning with Latent Alignment
Tongfan Guan (Chinese University of Hong Kong), Yun-Hui Liu (Chinese University of Hong Kong)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerOptical FlowImage
🎯 What it does: This paper proposes BridgeDepth, a unified framework that synchronizes monocular contextual information with binocular geometric matching through bidirectional cross-attention at the implicit representation layer, achieving simultaneous output of metric disparity and relative depth.
Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation
Bozhong Zheng (ShanghaiTech University), Yingna Wu (ShanghaiTech University)
Anomaly DetectionPoint Cloud
🎯 What it does: This paper studies an unsupervised 3D anomaly detection and repair framework based on pose-aware Signed Distance Function (SDF) called PASDF.
Bridging Class Imbalance and Partial Labeling via Spectral-Balanced Energy Propagation for Skeleton-based Action Recognition
Yandan Wang (North China Electric Power University), Weiming Dong (North China Electric Power University)
ClassificationRecognitionGraph Neural NetworkContrastive LearningVideoTime Series
🎯 What it does: This paper proposes a Spectral Energy Propagation with Balance (SpeLER) framework to address the issues of class imbalance and the coexistence of partial labels in skeleton action recognition.
Bridging Continuous and Discrete Tokens for Autoregressive Visual Generation
Yuqing Wang (University of Hong Kong), Xihui Liu (University of Hong Kong)
GenerationTransformerAuto EncoderImage
🎯 What it does: The TokenBridge method is proposed, which converts pre-trained continuous VAE features into discrete tokens through post-training dimension-level quantization, and based on this, achieves efficient generation under a large vocabulary using a dimension-level autoregressive head.
Bridging Diffusion Models and 3D Representations: A 3D Consistent Super-Resolution Framework
Yi-Ting Chen (University of Maryland), Jia-Bin Huang (University of Maryland)
RestorationSuper ResolutionDiffusion modelGaussian SplattingImage
🎯 What it does: A super-resolution framework based on 3D Gaussian Splatting, called 3DSR, is proposed, which combines diffusion models for multi-view consistency training.
Bridging Domain Generalization to Multimodal Domain Generalization via Unified Representations
Hai Huang (Zhejiang University), Zhou Zhao (Zhejiang University)
Domain AdaptationContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes a Unified Representation Framework (URMMDG) that maps the shared semantics of different modalities into the same space through supervised contrastive learning and information decoupling, and synchronously applies unimodal domain generalization (DG) techniques (such as Mixup, JiGen, IBN-Net) in this unified space, significantly enhancing multimodal domain generalization performance.
Bridging Local Inductive Bias and Long-Range Dependencies with Pixel-Mamba for End-to-end Whole Slide Image Analysis
Zhongwei Qiu (Zhejiang University), Le Lu (DAMO Academy Alibaba Group)
ClassificationSegmentationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Designed and implemented an end-to-end, memory-efficient whole slide image (WSI) analysis framework called Pixel-Mamba, capable of directly modeling whole slides from pixel-level without the need for pathology-specific pre-training, and completing tasks such as staging and prognosis.
Bridging the Gap between Brain and Machine in Interpreting Visual Semantics: Towards Self-adaptive Brain-to-Text Decoding
Jiaxuan Chen (Zhejiang University), Gang Pan (Zhejiang University)
RecognitionGenerationTransformerLarge Language ModelVision Language ModelImageTextMagnetic Resonance Imaging
🎯 What it does: An adaptive semantic decoding method called Mind-SA is proposed, which dynamically detects the image regions of interest in the brain and uses them as supervision to enhance the quality of brain-to-text reconstruction.
Bridging the Gap Between Ideal and Real-world Evaluation: Benchmarking AI-Generated Image Detection in Challenging Scenarios
Chunxiao Li (Beijing Normal University), Yao Zhu (Tsinghua University)
Object DetectionVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: RRDataset and RRBench are proposed to evaluate the robustness of AI-generated image detection models under real network propagation and re-digitization conditions, and the largest artificial detection benchmark experiment has been conducted.
Bridging the Skeleton-Text Modality Gap: Diffusion-Powered Modality Alignment for Zero-shot Skeleton-based Action Recognition
Jeonghyeok Do (Korea Advanced Institute of Science and Technology), Munchurl Kim (Korea Advanced Institute of Science and Technology)
RecognitionGraph Neural NetworkDiffusion modelContrastive LearningVideoMultimodality
🎯 What it does: A zero-shot skeleton action recognition framework based on diffusion models (TDSM) is proposed, which conditions text prompts during the reverse diffusion process to implicitly align skeleton features with text features, thereby achieving a unified latent space matching between skeletons and text.
Bridging the Sky and Ground: Towards View-Invariant Feature Learning for Aerial-Ground Person Re-Identification
Wajahat Khalid (University of Science and Technology of China), Muhammad Sher Afgan (University of Science and Technology of China)
RecognitionRetrievalTransformerImage
🎯 What it does: A cross-platform human weight recognition framework VIF-AGReID is proposed, aimed at learning view-invariant features.
Bring Your Rear Cameras for Egocentric 3D Human Pose Estimation
Hiroyasu Akada (Max Planck Institute for Informatics), Christian Theobalt (Max Planck Institute for Informatics)
Pose EstimationTransformerImage
🎯 What it does: Introducing a rear camera in an egocentric environment for full-body 3D pose estimation, and proposing a Transformer-based heatmap refinement module to improve estimation accuracy.
Bringing RNNs Back to Efficient Open-Ended Video Understanding
Weili Xu (Zhejiang University), Gaoang Wang (Zhejiang University)
Recurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodality
🎯 What it does: AURORALONG is proposed, replacing the Transformer LLM with a linear RNN RWKV, and combining visual token merging to achieve low memory efficient inference for long videos.
BUFFER-X: Towards Zero-Shot Point Cloud Registration in Diverse Scenes
Minkyun Seo (Seoul National University), Jaesik Park (Seoul National University)
Pose EstimationOptimizationSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: A completely zero-shot point cloud registration pipeline called BUFFER-X is proposed, which can achieve global registration in diverse scenarios without any manual parameter tuning or training data transfer.
BVINet: Unlocking Blind Video Inpainting with Zero Annotations
Zhiliang Wu (Zhejiang University), Yi Yang (Zhejiang University)
RestorationTransformerVideo
🎯 What it does: Proposed a blind video inpainting task and designed an end-to-end mask prediction network and video completion network to achieve this task.
C2MIL: Synchronizing Semantic and Topological Causalities in Multiple Instance Learning for Robust and Interpretable Survival Analysis
Min Cen (University of Science and Technology of China), Liansheng Wang (Xiamen University)
ClassificationExplainability and InterpretabilityGraph Neural NetworkTransformerContrastive LearningBiomedical Data
🎯 What it does: A dual causal graph-based multi-instance learning framework, C MIL, is proposed to eliminate semantic bias in WSI and identify causal subgraphs to enhance survival analysis.