CVPR 2026 Papers — Page 4
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Beyond Global Similarity: Multi-Conditional Retrieval for Fine-Grained Cross-Modal Understanding
Xuan Lu (Shanghai Jiao Tong University), Xiaoyu Shen (Eastern Institute of Technology)
RetrievalVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposed the MCMR (Multi-Conditional Multimodal Retrieval) benchmark to study multi-condition cross-modal retrieval under natural language queries;
Beyond Graph Model: Reliable VLM Fine-Tuning via Random Graph Adapter
Bo Jiang (Anhui University), Bin Luo (Anhui University)
ClassificationDomain AdaptationGraph Neural NetworkLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: Designed a vision-language model adapter based on a random graph, VRGAdapter, leveraging diverse text descriptions to construct probability distribution nodes and capturing intra-class diversity and inter-class relationships through graph convolution propagation.
Beyond Ground-Truth: Leveraging Image Quality Priors for Real-World Image Restoration
Fengyang Xiao (Duke University), Sina Farsiu (Duke University)
RestorationTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a framework called IQPIR, which leverages no-reference image quality assessment (NR-IQA) priors to guide real-world image restoration, aiming to address the issue of limited restoration quality caused by inconsistent ground-truth (GT) image quality in training data.
Beyond Heuristic Prompting: A Concept-Guided Bayesian Framework for Zero-Shot Image Recognition
Hui Liu (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
ClassificationRecognitionLarge Language ModelVision Language ModelImage
🎯 What it does: Propose a concept-based Bayesian framework (CGBC) that leverages large language models (LLMs) to generate class-specific concepts and adjusts their weights through an adaptive soft clipping likelihood function, achieving zero-shot image classification.
Beyond Layer-Wise Merging: Chain-of-Merging for Vision-Language Models
Xinyu Zhang (Ministry Of Education Laboratory Of Intelligent Networks And Network Security), Jun Liu (Ministry Of Education Laboratory Of Intelligent Networks And Network Security)
TransformerLarge Language ModelMultimodality
🎯 What it does: Propose the Chain-of-Merging (CoM) framework to enhance parameter fusion between Vision-Language Models (VLM) and Large Language Models (LLM), thereby improving visual and mathematical reasoning capabilities
Beyond Matching to Tiles: Bridging Unaligned Aerial and Satellite Views for Vision-Only UAV Navigation
Kejia Liu (Zhejiang University), Haofei Zhang (Zhejiang University)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: Propose a purely visual UAV navigation method called Bearing-UAV, which can simultaneously regress absolute position and heading angle under conditions where the UAV perspective and satellite perspective are misaligned.
Beyond Mimicry: Learning Whole-Body Human-Humanoid Interaction from Human-Human Demonstrations
Wei-Jin Huang (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Data SynthesisOptimizationRobotic IntelligenceTransformerDiffusion modelVideoMultimodalityMesh
🎯 What it does: This paper proposes a physical consistency transformation pipeline (PAIR) for generating human-robot interaction (HHoI) data from human-human interaction (HHI) data, and trains a disentangled spatiotemporal action reasoning network (D-STAR) based on this data to achieve full-body collaborative interaction.
Beyond Missing Modalities: Hypergraph Conditioned Diffusion for Uncertainty-Aware Multimodal Emotion Recognition
Xihang Qiu (Shenzhen MSU-BIT University), Chun Li (Shenzhen MSU-BIT University)
ClassificationRecognitionGraph Neural NetworkDiffusion modelMultimodality
🎯 What it does: For the emotion recognition task in dialogues, missing modalities are recovered and emotions are predicted through hypergraph diffusion and evidence fusion when modalities are missing.
Beyond Multiple Choice: Verifiable OpenQA for Robust Vision-Language RFT
Yesheng Liu (Institute of Automation, Chinese Academy of Sciences), Jiajun Zhang (Beijing Academy of Artificial Intelligence)
Large Language ModelReinforcement LearningPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Rewrite multiple-choice questions into verifiable open-ended questions to reduce the model's dependence on options.
Beyond Myopic Alignment: Lookahead Optimization for Online Class-Incremental Learning
Song Lai (City University of Hong Kong), Gaofeng Meng (Chinese Academy of Sciences)
OptimizationMeta LearningConvolutional Neural NetworkImageSequential
🎯 What it does: Propose a Lookahead Optimization for Rehearsal (LOR) method for online class-incremental learning, which explores multiple future model states under the current parameter setting by combining different weights of new task gradients and replay gradients, and guides updates using a soft worst-case (Log-Sum-Exp) loss aggregation, significantly reducing gradient conflicts and enhancing memory retention for old tasks.
Beyond Objects: Contextual Synthetic Data Generation for Fine-Grained Classification
William Yang (Princeton University), Olga Russakovsky (Princeton University)
ClassificationData SynthesisTransformerSupervised Fine-TuningVision Language ModelDiffusion modelImageText
🎯 What it does: Propose the BOB (Beyond Objects) strategy, which leverages a small number of real samples to extract class-agnostic attributes (background, pose). These attributes are incorporated as conditions during T2I model fine-tuning, and marginalization of attributes is applied during generation to produce synthetic data with greater diversity and better alignment with the target distribution, used for low-shot training in fine-grained classification.
Beyond Patches: Global-aware Autoregressive Model for Multimodal Few-Shot Font Generation
Haonan Cai (Peking University), Zhouhui Lian (Peking University)
GenerationTransformerReinforcement LearningMultimodality
🎯 What it does: Proposes GAR-Font, a few-shot font generation framework that combines a global-aware tokenizer and an autoregressive generator, supporting dual-modal input of visual and text.
Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs
Jingze Wu (Sun Yat-sen University), Hongbo Chen (Sun Yat-sen University)
OptimizationExplainability and InterpretabilityComputational EfficiencySupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodality
🎯 What it does: Proposed the VideoThinker framework for lightweight multimodal language model (MLLM) video reasoning, addressing the issue of models easily falling into perceptual shortcuts and limited reasoning capabilities during reinforcement learning fine-tuning.
Beyond Pixel Simulation: Pathology Image Generation via Diagnostic Semantic Tokens and Prototype Control
Minghao Han (Fudan University), Lihua Zhang (Fudan University)
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelFlow-based ModelImageTextMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Proposes UNIPATH, a framework that unifies diagnostic-level pathologic multimodal understanding models with controllable image generators, achieving fine-grained generation at text, semantic, and prototype levels through multi-stream control.
Beyond Prompt Degradation: Prototype-guided Dual-pool Prompting for Incremental Object Detection
Yaoteng Zhang (Northwestern Polytechnical University), Qi Wang (Northwestern Polytechnical University)
Object DetectionPrompt EngineeringImage
🎯 What it does: This paper addresses the prompt decay problem in incremental object detection by proposing the PDP framework.
Beyond Reassembly: Fractured Object Recovery with Missing Parts
Qun-Ce Xu (Tsinghua University), Shi-min Hu
RestorationPose EstimationConvolutional Neural NetworkTransformerAuto EncoderPoint CloudBenchmark
🎯 What it does: Studied a learning-based fragmented object recovery task that both reconstructs fragments and predicts missing parts.
Beyond Rule-Based Agents: Active Markov Games for Realistic Multi-Agent Interaction in Autonomous Driving
Yuan Gui (Northeastern University), Liqi Qu (Northeastern University)
Autonomous DrivingTransformerReinforcement LearningMultimodality
🎯 What it does: By constructing an Active Markov Game (AMG) framework and proposing a multi-agent co-evolution training mechanism, the study enhances the decision-making capabilities of autonomous driving systems in complex multi-vehicle interaction scenarios such as unsignalized intersections.
Beyond Scanpaths: Graph-Based Gaze Simulation in Dynamic Scenes
Luke Palmer (GlimpseML), Hazem Abdelkawy (Toyota Motor Europe)
Autonomous DrivingGraph Neural NetworkTransformerVideoGraphSequentialBenchmark
🎯 What it does: This paper proposes a graph-based dynamic gaze simulation framework, modeling driving scenes and driver gaze as spatiotemporal heterogeneous graphs. It predicts continuous raw gaze trajectories using Affinity Relation Transformer (ART) and Object Density Network (ODN), and releases a new Focus100 driving gaze dataset.
Beyond Semantic Search: Towards Referential Anchoring in Composed Image Retrieval
Yuxin Yang (CASIA), Weiming Hu (CASIA)
RetrievalTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes the Object Anchor Composite Image Retrieval (OACIR) task, which requires retrieving images while maintaining semantic matching and strictly preserving specified instances, and constructs a large-scale multi-domain annotated dataset called OACIRR;
Beyond Sequential Tools: A Unified VLM Agent System for Photographic Post-Processing via Dynamic Multi-Expert Fusion
Honglin Xiong (ShanghaiTech University), Qian Wang (ShanghaiTech University)
RestorationTransformerMixture of ExpertsVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes a unified photo post-processing framework led by a vision-language model, achieving dynamic multi-expert fusion to resolve multiple coupled distortions in one go.
Beyond Single Images: A Comprehensive Benchmark for Album-Level Vision-Language Understanding
Shawn Huang (Brigham Young University), Bryan Morse (Brigham Young University)
Prompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposes AlbumBench as a multi-image understanding benchmark for personal photo albums, defining four tasks: intent selection, intent scoring, group labeling, and group clustering.
Beyond Single Solution: Multi-Hypothesis Deep Unfolding Network for Image Compressive Sensing
Wenxue Cui (Harbin Institute of Technology), Debin Zhao (Harbin Institute of Technology)
RestorationConvolutional Neural NetworkTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed a multi-hypothesis collaborative deep unfolding network (MHC-DUN) for image compressive sensing reconstruction.
Beyond Single-View Sufficiency: CVBench for Cross-View Human Understanding
Tianchen Guo (University of Queensland), Xin Yu (Follow Me Ai Pty Ltd)
TransformerLarge Language ModelPrompt EngineeringVideoMultimodalityBenchmark
🎯 What it does: Proposes CVBench, a cross-perspective human understanding benchmark containing 3000 question-answer pairs requiring multi-perspective fusion.
Beyond Soft Label: Dataset Distillation via Orthogonal Gradient Matching
Deyu Bo (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Investigated methods for dataset distillation (DD) using a minimal number of synthetic samples on ImageNet-1K, proposing the Orthogonal Gradient Matching (OGM) framework.
Beyond Static Frames: Temporal Aggregate-and-Restore Vision Transformer for Human Pose Estimation
Hongwei Fang (Zhejiang Gongshang University), Wenwu Yang (Zhejiang Gongshang University)
Pose EstimationTransformerVideo
🎯 What it does: Proposed a video-based 2D human pose estimation framework named TAR-ViTPose, which can incorporate temporal modeling while maintaining the pure Transformer architecture of ViTPose.
Beyond Strict Pairing: Arbitrarily Paired Training for High-Performance Infrared and Visible Image Fusion
Yanglin Deng (Jiangnan University), Josef Kittler (University of Surrey)
Image HarmonizationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImageMultimodality
🎯 What it does: Propose a pairing training method for infrared and visible images without relying on strict alignment, aiming for high-performance fusion.
Beyond Success: Refining Elegant Robot Manipulation from Mixed-Quality Data via Just-in-Time Intervention
Yanbo Mao (Jilin University), Meibao Yao (Jilin University)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelFlow-based ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a method to significantly enhance the elegance and success rate of robotic manipulation by introducing real-time evaluation of execution quality and selective intervention during inference, based on a Vision-Language-Action model trained on mixed quality data.
Beyond Text Prompts: Precise Concept Erasure through Text-Image Collaboration
Jun Li, Guo-Sen Xie (Nanjing University Of Information Science And Technology)
GenerationTransformerVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: To address the issue of concept misuse in text-to-image diffusion models, this paper proposes a text-image collaborative concept erasure framework (TICoE), which can precisely remove specified concepts while maintaining generation quality.
Beyond Text: Visual Description Assembly by Probabilistic Model for CLIP-based Weakly Supervised Semantic Segmentation
Xianglin Qiu (XJTLU), Jimin Xiao (XJTLU)
SegmentationVision Language ModelFlow-based ModelContrastive LearningImage
🎯 What it does: Construct instance-specific visual description prototypes using CLIP visual features to generate more complete class activation maps (CAM) and improve the accuracy of weakly supervised semantic segmentation.
Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations
Tuan Dung Nguyen (Hanoi University of Science and Technology), Vu Minh Hieu Phan (University of Adelaide)
Anomaly DetectionExplainability and InterpretabilityTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Propose a token-level hallucination detection framework based on fine-grained visual attention distribution and cross-modal alignment consistency, utilizing Attention Dispersion Score (ADS) and Cross-Modal Grounding Consistency (CGC) to discriminate the generated text of LVLMs.
Beyond the Golden Data: Resolving the Motion-Vision Quality Dilemma via Timestep Selective Training
Xiangyang Luo (Tsinghua University), Shao-Lun Huang (Tsinghua University)
GenerationDiffusion modelVideo
🎯 What it does: This paper investigates the negative correlation between visual quality and motion quality in video generation models, proposing a Temporal Quality Disentanglement (TQD) method that performs gradient hierarchical selection training on imbalanced quality data, demonstrating that using single-dimensional quality data can achieve or even surpass the training effects of traditional golden data.
Beyond the Ground Truth: Enhanced Supervision for Image Restoration
Donghun Ryou (Seoul National University), Bohyung Han (Seoul National University)
RestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Propose a supervised enhancement framework that generates multi-scale perceptually enhanced ground truth images via a first-order diffusion super-resolution model. Then, an adaptive frequency mask generator is used in the frequency domain to blend the original ground truth with its super-resolution variant, resulting in higher quality enhanced ground truth. Subsequently, a lightweight output refinement network (ORNet) is trained and can be seamlessly integrated into any existing image restoration model to further improve perceptual quality.
Beyond the Static World: Continual Category Discovery under Visual Drift
Wei Feng (Monash University), Zongyuan Ge (Monash University)
ClassificationDomain AdaptationTransformerImage
🎯 What it does: Studied how to automatically discover known and unknown categories in a streaming unlabeled data scenario under open continuous category discovery (OCCD), without access to labeled data and facing domain drift.
Beyond the Static-World: Lifelong Learning for All-in-One Medical Image Restoration
Shihao Shan (Tianjin University), Jingjing Deng (University of Bristol)
RestorationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: Proposed a lifelong learning framework called ROME for全域 medical image restoration, aiming to simultaneously handle the restoration tasks of multi-modal images (MRI, CT, PET) and retain memory of historical knowledge in dynamic data streams.
Beyond Tie Points: Satellite Image Block Adjustment based on Dense Feature Consistency
Yi Liu (Wuhan University), Yongjun Zhang (Wuhan University)
OptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a satellite image block adjustment method that does not rely on sparse matching points, directly optimizing global correction parameters through dense feature consistency.
Beyond Top Activations: Efficient and Reliable Crowdsourced Evaluation of Automated Interpretability
Tuomas Oikarinen (UC San Diego), Tsui-Wei Weng (UC San Diego)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes two methods (Model-Guided Importance Sampling and Bayes Rating Aggregation) for low-cost, reliable human evaluation of neural explanations, and compares multiple automatic explanation methods in large-scale experiments.
Beyond Weak Supervision: MLLMs-Guided Graded Knowledge Distillation for Unsupervised Camouflaged Object Detection
Huafeng Chen (Nanjing University), Caifeng Shan (Nanjing University)
Object DetectionKnowledge DistillationTransformerLarge Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose an unsupervised camouflaged object detection (UCOD) method named UCOD-MKD, leveraging multi-modal large language models (MLLM) and the Segment Anything Model (SAM) to generate high-quality pseudo-labels, and enhancing student model performance through graded knowledge distillation (GKD).
Beyond What's Shared: Recovering Lost Unique Information from Intermediate Layers to Boost Multimodal Geo-Foundation Models
JangHyeon Lee (University of Minnesota), Dalton Lunga (Oak Ridge National Laboratory)
Representation LearningContrastive LearningImageMultimodalityTabular
🎯 What it does: Propose a hierarchical fusion method (BWS) without additional objectives or external models, improving multimodal representations of geographical locations and satellite imagery by analyzing unique information retained in intermediate layers of multimodal contrastive learning models;
Bezier Degradation Modeling for LiDAR-based Human Motion Capture
Xiaoqi An (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Pose EstimationConvolutional Neural NetworkGraph Neural NetworkTransformerPoint Cloud
🎯 What it does: This paper proposes a hierarchical motion representation and reduction strategy based on B’-Bezier curves, combining time-scale motion Transformer and multi-layer motion aggregator to reconstruct 3D human motion from sparse LiDAR point clouds in a coarse-to-fine manner.
BHCast: Unlocking Black Hole Plasma Dynamics from a Single Blurry Image with Long-Term Forecasting
Renbo Tu (University of Toronto), Aviad Levis (University of Toronto)
Convolutional Neural NetworkImageVideoPhysics Related
🎯 What it does: Predict the subsequent high-resolution temporal evolution of blurred black hole images from a single frame, and extract physical features to infer black hole spin and inclination.
Bi-Bridge: Bidirectional Diffusion Bridges for Low-Light Image Enhancement
Zeyu Hua (Northeast Normal University), Caixia Zheng (Northeast Normal University)
RestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes the Bi-Bridge framework, which introduces a bidirectional diffusion bridge in low-light image enhancement tasks to enforce the network to learn symmetric relationships between enhancement and degradation, thereby improving content fidelity.
Bi-directional Autoregressive Diffusion for Large Complex Motion Interpolation
Yongrui Ma (ByteDance Inc.), Tianfan Xue (CUHK MMLab)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Propose a high-quality video frame interpolation method called ARVFI based on a bidirectional autoregressive diffusion model, which first estimates intermediate motion in the DINOv3 feature space and then generates frames conditionally;
Bias at the End of the Score
Salma Abdel Magid (Princeton University), Olga Russakovsky (Princeton University)
GenerationOptimizationExplainability and InterpretabilityImageMultimodality
🎯 What it does: This study systematically evaluates the bias and robustness of reward models in text-to-image generation, revealing significant unfair impacts on gender and race during the optimization process.
Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models
Ivan Luiz De Moura Matos (Institut Polytechnique de Paris), Enzo Tartaglione (Institut Polytechnique de Paris)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImageText
🎯 What it does: In traditionally trained deep networks, structured pruning and learnable masks are used to directly extract bias-invariant subnetworks from within the model without requiring retraining or additional bias-invariant data.
Bias Is a Subspace, Not a Coordinate: A Geometric Rethinking of Post-hoc Debiasing in Vision-Language Models
Dachuan Zhao (Harvard University), Yongchao Chen (MIT)
Explainability and InterpretabilityVision Language ModelMultimodality
🎯 What it does: Proposed a post-hoc debiasing framework called Subspace Projection Debiasing (SPD), which identifies and projects away the linear bias subspace in VLM embeddings, followed by inserting a neutral mean to preserve semantic integrity.
Bidirectional Cross-Modal Prompting for Event-Frame Asymmetric Stereo
Ninghui Xu (Southeast University), Stefano Mattoccia (University of Bologna)
Depth EstimationConvolutional Neural NetworkPrompt EngineeringMultimodality
🎯 What it does: Propose the Bi-CMPStereo bidirectional cross-modal prompting framework, which uses event-frame heterogeneous stereo vision to achieve high-precision disparity estimation
Bidirectional Multimodal Prompt Learning with Scale-Aware Training for Few-Shot Multi-Class Anomaly Detection
Yujin Lee (Yonsei University), Hyunsoo Yoon (Yonsei University)
Anomaly DetectionTransformerPrompt EngineeringVision Language ModelMultimodalityBiomedical Data
🎯 What it does: Propose a lightweight multimodal prompt learning framework called AnoPLe to address the problem of few-shot multi-class anomaly detection;
Bidirectional Normalizing Flow: From Data to Noise and Back
Yiyang Lu (MIT), Kaiming He (MIT)
GenerationTransformerFlow-based ModelImage
🎯 What it does: Propose Bidirectional Normalizing Flow (BiFlow), which introduces a learnable inverse model while maintaining the reversibility of forward NF, eliminating the limitations of explicit inverse, achieving high-quality generation in a single step.
Bidirectional Query-Driven Generation of Parametric CAD Sketch
Yang Liu (Beijing Institute of Technology), Fang Deng (Beijing Institute of Technology)
GenerationTransformerGraph
🎯 What it does: Propose a CADSketcher framework based on bidirectional queries for automatically generating complete executable parameterized CAD sketches from partial sketches with arbitrary spans.
BiEvLight: Bi-level Learning of Task-Aware Event Refinement for Low-Light Image Enhancement
Zishu Yao (Fuzhou University), Xing Chen (Fuzhou University)
RestorationOptimizationConvolutional Neural NetworkMultimodality
🎯 What it does: Propose BiEvLight, a task-aware framework based on two-layer optimization, which jointly trains event denoising and low-light image enhancement to achieve complementary fusion of event information and image features.
BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation
Yasong Dai (Australian National University), Hongdong Li (Australian National University)
GenerationSupervised Fine-TuningFlow-based ModelImage
🎯 What it does: This paper proposes a bidirectional flow matching framework called BiFM, which can simultaneously learn image generation and inversion under few-step sampling conditions.
BiGain: Unified Token Compression for Joint Generation and Classification
Jiacheng Liu (VILA Lab, Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (VILA Lab, Mohamed bin Zayed University of Artificial Intelligence)
ClassificationGenerationTransformerDiffusion modelImage
🎯 What it does: To address the acceleration problem of diffusion models, the BiGain framework is proposed, which can significantly improve the classification performance of the same model while maintaining generation quality.
BiGMINT: Biologically-guided Hierarchical Multimodal Integration for Modeling Multiple Compound Activities in Drug Discovery
Pushpak Pati (Janssen R&D, LLC), Zhoubing Xu (Janssen R&D, LLC)
Drug DiscoveryTransformerBiomedical Data
🎯 What it does: Develop the BiGMINT framework, which hierarchically integrates chemoproteomics with high-content imaging (HCI) data, using chemoproteomics-guided phenotypic feature aggregation, task-aware cross-modal fusion, and protein-protein interaction (PPI) priors to predict multi-compound activities.
Bilevel Layer-Positioning LoRA for Real Image Dehazing
Yan Zhang (Sun Yat-sen University), Zhuo Su (Sun Yat-sen University)
RestorationDomain AdaptationVision Language ModelImage
🎯 What it does: Propose an unsupervised semantic guidance loss based on CLIP (H2C) and a two-level optimization strategy for LoRA layer localization (BiLaLoRA), for efficient image dehazing from synthetic domains to real fog scenes;
BiMotion: B-spline Motion for Text-guided Dynamic 3D Character Generation
Miaowei Wang (University Of Edinburgh), Amir Vaxman (University Of Edinburgh)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelAuto EncoderTextMesh
🎯 What it does: Propose the BiMotion framework, utilizing B-spline curves to achieve continuous, controllable text-driven 3D character animation generation.
BinaryAttention: One-Bit QK-Attention for Vision and Diffusion Transformers
Chaodong Xiao (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
ClassificationSegmentationGenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Propose a one-bit QK-attention model named BinaryAttention, which computes attention using binary queries and keys, significantly improving speed while maintaining or even surpassing full-precision performance.
BiomedCCPL: Causal Conditional Prompt Learning for Biomedical Vision-Language Models
Xueliang Cui (Southern University of Science and Technology), Ruxin Wang (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
Prompt EngineeringVision Language ModelBiomedical Data
🎯 What it does: Propose BiomedCCPL, a framework based on causal conditional prompt learning, designed for efficient adaptation of medical vision-language models in few-shot scenarios.
BiOTPrompt: Bidirectional Optimal Transport Guided Prompting for Disease Evolution-aware Radiology Report Generation
Tengfei Liu (Beijing University of Technology), Baocai Yin (University of Science and Technology of China)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBiomedical DataElectronic Health Records
🎯 What it does: Propose the BiOTPrompt framework, which utilizes bidirectional optimal transport to identify proliferative/regressive regions in historical and current chest X-rays, and constructs dynamic prompts to guide LLMs in generating progress-aware, factually accurate radiology reports.
BioVITA: Biological Dataset, Model, and Benchmark for Visual-Textual-Acoustic Alignment
Risa Shinoda (University of Osaka), Fumio Okura (University of Osaka)
RetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataBenchmarkAudio
🎯 What it does: Proposed the BioVITA framework, which includes a million-scale audio-image-text training set, a unified representation model, and a cross-modal retrieval benchmark.
BiPA: Bilevel Prompt Adaptation for Underwater Instance Segmentation
Long Ma (Dalian University of Technology), Weimin Wang (Dalian University of Technology)
SegmentationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerPrompt EngineeringImage
🎯 What it does: In the underwater instance segmentation task, this paper proposes the BiPA method, which utilizes dual-layer optimization to learn specialized dense prompts and effectively transfers SAM to the underwater domain through dual-modal prompts (dense + sparse) and a foreground attention injection module (FAI).
BiPreManip: Learning Affordance-Based Bimanual Preparatory Manipulation through Anticipatory Collaboration
Yan Shen (Peking University), Hao Dong (Peking University)
Pose EstimationRobotic IntelligenceConvolutional Neural NetworkAuto EncoderTextPoint Cloud
🎯 What it does: This paper proposes the BiPreManip framework, which uses vision-based affordance anticipation to predict the target arm's final interaction, guiding the assisting arm to perform preparatory actions and creating suitable conditions for the target arm's subsequent operations, achieving cooperative pre-processing manipulation with dual arms;
BiProLoRA: Bilevel Prompt LoRA for Real Scene Recovery
Nan An (Dalian University of Technology), Risheng Liu (Dalian University of Technology)
RestorationDomain AdaptationOptimizationHyperparameter SearchPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: For real-world restoration tasks, we propose BiProLoRA: an adaptation paradigm that combines self-supervised distribution consistency learning (DFL) with a two-layer Prompt-LoRA to achieve high-fidelity restoration from synthetic data to real data.
BIT: Matching-based Bi-directional Interaction Transformation Network for Visible-Infrared Person Re-Identification
Haoxuan Xu (Beihang University), Guanglin Niu (Beihang University)
RecognitionRetrievalTransformerContrastive LearningImage
🎯 What it does: Propose a matching-based Bidirectional Interaction Transformation Network (BIT), addressing the modality gap and data imbalance in visible-infrared person re-identification through bidirectional interaction and adaptive local matching between visible and infrared image pairs.
Black-Box Domain Adaptation for Object Detection with Retention-Driven Knowledge Compression
Yuwu Lu (South China Normal University), Chunzhi Liu (South China Normal University)
Object DetectionDomain AdaptationKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: In the black-box domain adaptation scenario, we propose Retention-Driven Knowledge Compression (RDKC), which promotes the transfer of object detectors under source-free data/model conditions through Memory Retention and Scene Compression.
Black-box Membership Inference Attacks on the Pre-training Data of Image-generation Models
Tao Qi (Beijing University of Posts and Telecommunications), Yongfeng Huang (Tsinghua University)
Safty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Propose the SD-MIA framework for black-box membership inference attacks targeting pre-training data of diffusion-based image generation models.
BlackMirror: Black-Box Backdoor Detection for Text-to-Image Models via Instruction-Response Deviation
Feiran Li (Institute of Information Engineering, CAS), Qingming Huang (School of Computer Science and Technology, University of Chinese Academy of Sciences)
Anomaly DetectionSafty and PrivacyLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a black-box backdoor detection framework called BlackMirror for text-to-image models, which can identify various backdoor attacks without accessing the model's internal structure.
Blink: Dynamic Visual Token Resolution for Enhanced Multimodal Understanding
Yuchen Feng (Institute of Information Engineering, Chinese Academy of Sciences), Haifeng Wang (Baidu Inc)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose the Blink framework to dynamically enhance the visual perception capabilities of multimodal large language models (MLLMs), simulating the human 'rapid blinking'-style visual scanning and focusing process;
Block-based Learned Image Compression without Blocking Artifacts
Jong Wook Kim (Kyung Hee University), Hui Yong Kim (Kyung Hee University)
CompressionConvolutional Neural NetworkImage
🎯 What it does: Propose a block-based learning image compression framework that does not require retraining, utilizing analytical formulas for overlap propagation of convolution and transposed convolution to calculate the minimum overlap, ensuring complete consistency between block processing and full-image inference.
Block-Sparse Global Attention for Efficient Multi-View Geometry Transformers
Chung-Shien Brian Wang (RWTH Aachen University), Bastian Leibe (RWTH Aachen University)
Pose EstimationDepth EstimationTransformerImagePoint Cloud
🎯 What it does: Proposed a training-agnostic, block-sparse global attention alternative to accelerate inference in multi-view geometry Transformers (VGGT, π³, MapAnything).
BluRef: Unsupervised Image Deblurring with Dense-Matching References
Bang-Dang Pham (University of Wisconsin-Madison), Minh Hoai (Qualcomm AI Research)
RestorationConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes a fully unsupervised image deblurring framework called BluRef, which utilizes dense matching to generate pseudo-real sharp images from unpaired blurry images and reference sharp images collected from the target domain, thereby iteratively training the deblurring network.
Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing
Cheng Cui (Baidu Inc.), Yanjun Ma (Baidu Inc.)
RecognitionComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose PaddleOCR-VL—a two-stage coarse-to-fine document parsing framework that first uses VRFM to lightweight locate effective visual regions and predict reading order, then employs a 0.9B vision-language model to perform fine-grained recognition within cropped regions, achieving efficient parsing at high resolution.
Boosting Quantitive and Spatial Awareness for Zero-Shot Object Counting
Da Zhang (Northwestern Polytechnical University), Junyu Gao (Northwestern Polytechnical University)
Object DetectionTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposes a new zero-shot object counting framework QICA, achieving accurate density estimation through quantity-aware collaborative prompts and cost aggregation decoder.
Boosting Reasoning in Large Multimodal Models via Activation Replay
Yun Xing (Nanyang Technological University), Yu-Gang Jiang (Fudan University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Analyze the internal mechanisms of large multimodal models (LMMs) post-trained with RLVR (Reinforcement Learning with Verifiable Rewards), and propose the Activation Replay method that does not require additional training. This method enhances reasoning capabilities by adding learnable noise to visual tokens during testing, making low-entropy activations of the RLVR model closer to those of the original LMM.
Boosting Self-Supervised Tracking with Contextual Prompts and Noise Learning
Yaozong Zheng (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
Object TrackingRepresentation LearningTransformerVideo
🎯 What it does: Designed a self-supervised visual tracking framework called PNTrack, which leverages a dual-modal context association mechanism to learn robust representations through prompts and noise during training.
Boosting Vision-Language Models Towards Cross-Domain Incremental Object Detection
Xu Wang (University of Science and Technology of China), Zilei Wang (University of Science and Technology of China)
Object DetectionDomain AdaptationImageBenchmark
🎯 What it does: Proposed the cross-domain incremental object detection (CDIOD) benchmark and the Dynamic Group Subspace (DGS) framework to achieve cross-domain incremental learning.
Boosting Vision-Language-Action Finetuning with Feasible Action Neighborhood Prior
Haochen Niu (Shanghai Jiao Tong University), Fei Wen (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageMultimodality
🎯 What it does: Propose a regularization method based on the Feasible Action Neighborhood (FAN) to guide Vision-Language-Action (VLA) models during fine-tuning to generate smoother, more fault-tolerant action distributions, thereby improving sampling efficiency and generalization capabilities.
Boosting Visual Reprogramming for CLIP with Dual Granularity Alignment
Jiayang Wu (Harbin Institute Of Technology (Shenzhen)), Weili Guan (Harbin Institute Of Technology (Shenzhen))
ClassificationTransformerContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Propose a Dual Granularity Alignment (DGA) framework for visual reprogramming in the CLIP model, combining visual hierarchy (multi-scale images) and semantic hierarchy (label hierarchy) to enhance classification performance in downstream tasks.
BoostSLT: Boosting Sign Language Translation via a Plug-and-Play Diffusion-Based Semantic Enhancer
Changzhou Han (Swinburne University of Technology), Yang Xiang (Swinburne University of Technology)
Large Language ModelDiffusion modelVideo
🎯 What it does: Propose BoostSLT, a pluggable sign language translation framework that employs unsupervised energy-aware segmentation and diffusion-based semantic reconstruction to enhance the coherence of long-sentence translation.
Bootstrap Dynamic-Aware 3D Visual Representation for Scalable Robot Learning
Qiwei Liang (Hong Kong University of Science and Technology (Guangzhou)), Renjing Xu (Hong Kong University of Science and Technology (Guangzhou))
Representation LearningRobotic IntelligenceTransformerDiffusion modelContrastive LearningWorld ModelPoint Cloud
🎯 What it does: Proposed an unsupervised 3D visual pre-training framework AFRO, capable of learning dynamic-aware 3D representations for robotic manipulation.
Bootstrap Your Own AV-Proxies: Adaptive Contrastive and Prototype Learning for Audio-Visual Segmentation
Junbo Zhang (Wuhan University), Chao Sun (Wuhan University)
SegmentationTransformerContrastive LearningVideoMultimodalityBenchmarkAudio
🎯 What it does: This study proposes a new audio-visual segmentation framework called BYOAVP, which adaptively aligns and denoises to address the cross-modal semantic gap and single-modal noise issues in audio-visual segmentation.
Bootstrapping Multi-view Learning for Test-time Noisy Correspondence
Changhao He (Sichuan University), Peng Hu (Sichuan University)
RecognitionRestorationDomain AdaptationRepresentation LearningImageTextMultimodality
🎯 What it does: This paper proposes a Bootstrapping-based multi-view learning framework, BML, to address the test-time view correspondence noise (TNC) problem. It directly generates noisy samples on the training set through adaptive masking and view shuffling, and employs a lightweight estimator to perform supervised learning on the reliability of each view, thereby achieving robust multi-view fusion.
Bootstrapping Video Semantic Segmentation Model via Distillation-assisted Test-Time Adaptation
Jihun Kim (KAIST), Kuk-Jin Yoon (KAIST)
SegmentationDomain AdaptationComputational EfficiencyKnowledge DistillationTransformerContrastive LearningVideo
🎯 What it does: Pre-trained image semantic segmentation models are converted into temporal semantic segmentation models that can adapt using only a few video frames through test-time distillation and lightweight attention fusion.
BOP-ASK: Object-Interaction Reasoning for Vision-Language Models
Vineet Bhat (New York University), Jonathan Tremblay (NVIDIA)
Data SynthesisPose EstimationSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed a large-scale BOP-Ask dataset for training and evaluating VLMs in object interaction reasoning, and provided two test benchmarks (BOP-Ask-core and BOP-Ask-lab).
Boundary-Responsive Differentiable Gating for Superpixel-Based Segmentation
Fatmaelzahraa Ahmed (Hamad Medical Corporation), Shidin Balakrishnan (Hamad Medical Corporation)
SegmentationConvolutional Neural NetworkContrastive LearningImageVideoBiomedical Data
🎯 What it does: Proposed the BRDG framework, which achieves efficient and accurate semantic segmentation in surgical scenes by utilizing differentiable superpixels and boundary response gating.
Breaking Multimodal LLM Safety via Video-Driven Prompting
Dong Wang (Hong Kong Polytechnic University), Bin Xiao (Hong Kong Polytechnic University)
Safty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: Study the safety of multimodal large language models (MLLMs) in video modalities and design video-based jailbreak attacks, proposing the Safety-Proximal Typographic Video (SPTV) method to bypass safety alignment.
Breaking Semantic Boundaries: Distribution-Guided Semantic Exploration for Creative Generation
Fu Feng (Southeast University), Xin Geng (Southeast University)
GenerationVision Language ModelDiffusion modelImageText
🎯 What it does: Propose the Distribution-Conditional Generation approach and the DisTok encoder-decoder framework, leveraging class distributions to guide the generation of novel visual concepts;
Breaking Smooth-Motion Assumptions: A UAV Benchmark for Multi-Object Tracking in Complex and Adverse Conditions
Jingtao Ye (Xidian University), Liang Zhang (Xidian University)
Object TrackingConvolutional Neural NetworkVideoBenchmark
🎯 What it does: Proposes DynUAV—a benchmark for multi-object tracking under complex high-speed maneuvers from a UAV perspective, containing 42 video clips, 1.7 million annotated frames, covering eight categories including vehicles, pedestrians, and industrial machinery.
Breaking Spurious Correlations: Uncertainty-Driven Causal Transformers for AU Detection
Yuru Wang (Northeast Normal University), Yue Zhou (Northeast Normal University)
RecognitionTransformerImage
🎯 What it does: Proposed a Transformer framework (UDCT) that integrates uncertainty modeling with causal intervention for facial action unit (AU) detection.
Breaking the 3D Dataset Bottleneck: Fast Scalable Generation of Aligned 3D Assets from Scratch for Category 6D Pose Estimation and Robotic Grasping
Duret Guillaume (Centrale Lyon), Liming Chen (Centrale Lyon)
GenerationData SynthesisPose EstimationRobotic IntelligenceTransformerLarge Language ModelDiffusion modelImageMesh
🎯 What it does: This paper proposes an end-to-end scalable pipeline that directly generates fully aligned 3D assets from text category prompts for 6D pose estimation and robot grasping.
Breaking the Continuum: Discrete Distribution Learning for Structural MRI Reconstruction
Tianle Lyu (Shenzhen Institutes of Advanced Technology), Yanjie Zhu (Shenzhen Institutes of Advanced Technology)
RestorationConvolutional Neural NetworkDiffusion modelScore-based ModelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a magnetic resonance imaging reconstruction framework called DiCoS, which combines discrete structure reasoning with continuous refinement to better capture discrete anatomical priors in medical images and improve reconstruction quality.
Breaking the Illusion: When Positive Meets Negative in Multimodal Decoding
Yubo Jiang (Beihang University), Haopeng Zhang (Beihang University)
Explainability and InterpretabilityVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This work proposes a no-training, plug-and-play dual-path decoding framework called PND, which dynamically balances visual likelihood and language prior during inference through positive and negative decoding paths, suppressing hallucinations in multimodal models.
Breaking the Regional Perception Bottleneck of Multimodal Large Language Models via External Reasoning Framework
Jinrong Zhang (Harbin Institute of Technology), Jianlong Wu (Shenzhen Loop Area Institute)
Object DetectionTransformerVision-Language-Action ModelContrastive LearningMultimodality
🎯 What it does: Investigate the bottleneck of multi-modal large language models (MLLMs) in pixel-level region perception (object localization) and propose the R-Ground framework, which significantly enhances localization performance by amplifying reasoning through multi-modal Monte Carlo Tree Search (MCTS) during the semantic refinement stage.
Breaking the Scalability Limit of Multi-Projector Calibration with Embedded Cameras
Takumi Kawano (University of Osaka), Daisuke Iwai (University of Osaka)
OptimizationComputational EfficiencyImage
🎯 What it does: In this study, the authors embedded multiple cameras on a calibration board to achieve simultaneous projection of structured light by multiple projectors and geometric calibration, thereby breaking the linear time bottleneck of traditional projector-by-projector calibration methods.
BrepGaussian: CAD reconstruction from Multi-View Images with Gaussian Splatting
Jiaxing Yu (Nanjing University), Yanwen Guo (Nanjing University)
SegmentationGenerationContrastive LearningGaussian SplattingImageMesh
🎯 What it does: Directly reconstructing CAD models (B-Rep representation) from multi-view images
BrepVGAE: Variational Graph Autoencoder with Unified Latent Representation for B-rep
Hao Guo (Northwestern Polytechnical University), Yilei Shi (Northwestern Polytechnical University)
GenerationRepresentation LearningGraph Neural NetworkAuto EncoderPoint CloudGraph
🎯 What it does: Propose the BrepVGAE framework, which uses a sparse graph variational autoencoder and a set decoder to uniformly encode and decode the geometry and topology of CAD B-rep models.
Brewing Stronger Features: Dual-Teacher Distillation for Multispectral Earth Observation
Filip Wolf (University of Ljubljana), Luka Čehovin Zajc (University of Ljubljana)
ClassificationSegmentationKnowledge DistillationRepresentation LearningTransformerContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Propose the dual-teacher distillation framework DEO, achieving unified representation learning on multispectral and optical inputs.
BrickNet: Graph-Backed Generative Brick Assembly
Peter Kulits (Inria), Cordelia Schmid (Inria)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: In this paper, we construct an autoregressive framework based on large language models for generating LEGO brick construction sequences and propose a graphical connection parameterization method.
Bridge: Basis-Driven Causal Inference Marries VFMs for Domain Generalization
Mingbo Hong (University of Twente), Hao Cheng (University of Twente)
Object DetectionDomain AdaptationKnowledge DistillationTransformerDiffusion modelImageBenchmark
🎯 What it does: Propose the Bridge framework, leveraging the front-door adjustment from causal inference and the Causal Basis Block (CBB) based on low-rank basis learning, to achieve cross-domain general object detection under single-source limited data by freezing Vision Foundation Models (VFMs);
BridgeEQA: Virtual Embodied Agents for Real Bridge Inspections
Subin Varghese (University of Houston), Vedhus Hoskere (University of Houston)
Robotic IntelligenceGraph Neural NetworkLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the BridgeEQA benchmark, which constructs an open-vocabulary embedded question-answering (EQA) dataset for real-world bridge inspection tasks, and designs the Embodied Memory Visual Reasoning (EMVR) method to address multi-perspective image retrieval and reasoning problems; meanwhile, it introduces Image Citation Relevance as a metric to measure the relevance of supporting images to answers.
Bridging Brain and Semantics: A Hierarchical Framework for Semantically Enhanced fMRI-to-Video Reconstruction
Yujie Wei (Fudan University), Hongming Shan (Ant Group)
GenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelVideoMultimodalityBiomedical Data
🎯 What it does: Propose a two-stage hierarchical framework called CINENEURON, which first maps fMRI to embeddings enriched with multimodal semantics (including text, images, actions, and categories) through bottom-up semantic enrichment, and then enhances the fMRI-to-video reconstruction quality by dynamically retrieving and fusing previous multimodal memories via a top-down memory fusion module called Mixture-of-Memories.