ECCV 2024 Papers — Page 3
European Conference on Computer Vision · 2387 papers
Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search
Lujun Li (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
GenerationNeural Architecture SearchGenerative Adversarial NetworkImageBenchmark
🎯 What it does: This paper proposes Auto-GAS, a fully training-free generative model architecture search framework, which automatically discovers and evolves 'zero-cost proxies' to rapidly evaluate candidate generators and ultimately searches for efficient GAN architectures.
AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion
Yitong Jiang (Chinese University of Hong Kong), Jinwei Gu (Chinese University of Hong Kong)
RestorationPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderImage
🎯 What it does: Propose AutoDIR, an automatic and universal image restoration system based on latent diffusion models, capable of automatically identifying and handling various unknown degradations, and supporting user personalization through text prompts.
AutoEval-Video: An Automatic Benchmark for Assessing Large Vision Language Models in Open-Ended Video Question Answering
Xiuyuan Chen (Shanghai Jiao Tong University), Weiran Huang (Shanghai Jiao Tong University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper creates the AutoEval-Video benchmark, constructing 327 instances containing videos, open-ended questions, and dedicated evaluation rules. It uses GPT-4 as an automatic evaluator to assess the performance of large vision-language models on open-ended video question-answering tasks.
Avatar Fingerprinting for Authorized Use of Synthetic Talking-Head Videos
Ekta Prashnani (NVIDIA), Orazio Gallo (NVIDIA)
RecognitionSafty and PrivacyConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: Investigated a method for verifying the identity of synthetic video avatars, termed Avatar Fingerprinting;
AvatarPose: Avatar-guided 3D Pose Estimation of Close Human Interaction from Sparse Multi-view Videos
Feichi Lu (Hong Kong University of Science and Technology), Otmar Hilliges
Pose EstimationNeural Radiance FieldVideo
🎯 What it does: Propose the AvatarPose method, which reconstructs personalized implicit neural avatars for each individual from sparse multi-view videos and uses them as priors for 3D pose and shape estimation.
AWOL: Analysis WithOut synthesis using Language
Silvia Zuffi (IMATI-CNR), Michael J. Black (Max Planck Institute for Intelligent Systems)
GenerationVision Language ModelFlow-based ModelImageTextMesh
🎯 What it does: Leverage language and CLIP latent space mapping to control existing 3D shape models (SMAL+ for animals and Tree-Gen for trees), generating unseen animal breeds and tree species with only a few labeled samples.
Background Adaptation with Residual Modeling for Exemplar-Free Class-Incremental Semantic Segmentation
Anqi Zhang (Beijing Institute of Technology), Guangyu Gao (Beijing Institute of Technology)
SegmentationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a background adaptation residual modeling method to address the background drift problem in sample-free incremental semantic segmentation.
Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding
Talfan Evans (Google DeepMind), Olivier Henaff (Google DeepMind)
ClassificationRetrievalComputational EfficiencyData-Centric LearningTransformerContrastive LearningImageMultimodality
🎯 What it does: Propose a data prioritization method based on active learning for large-scale visual pre-training (ClassAct and ActiveCLIP), which significantly reduces training iterations and FLOPs by using a small proxy model to compute the 'learnability' score of data.
BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting
Lingzhe Zhao (Westlake University), Peidong Liu (Westlake University)
RestorationPose EstimationOptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingImagePoint Cloud
🎯 What it does: By jointly learning the camera trajectory during exposure and 3D Gaussian point cloud parameters, a virtual sharp image is generated and averaged using a physical motion blur model, ultimately achieving motion blur removal and high-quality 3D scene reconstruction with real-time rendering support.
BAFFLE: A Baseline of Backpropagation-Free Federated Learning
Haozhe Feng (State Key Lab of CAD&CG, Zhejiang University), Min Lin (Sea AI Lab)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: Propose the BAFFLE framework, which replaces backpropagation in federated learning with multiple forward inferences and zeroth-order optimization, achieving gradient-free federated learning;
BAGS: Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling
Cheng Peng (Johns Hopkins University), Rama Chellappa (Johns Hopkins University)
RestorationGenerationConvolutional Neural NetworkNeural Radiance FieldGaussian SplattingImagePoint Cloud
🎯 What it does: Propose the BAGS method, integrating a Blur Proposal Network and multi-scale convolutional kernels into the 3D Gaussian Splatting framework to achieve robust reconstruction and novel view synthesis for real-world image blurs, including motion blur, defocus blur, and resolution inconsistency.
BAM-DETR: Boundary-Aligned Moment Detection Transformer for Temporal Sentence Grounding in Videos
Pilhyeon Lee (Inha University), Hyeran Byun (Yonsei University)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: Proposed a dual-channel decoding framework called BAM-DETR based on Transformer, which predicts time segments in videos that match natural language descriptions using boundary-aligned triplets (anchor point, start-end distances), and introduces ranking based on localization quality.
BAMM: Bidirectional Autoregressive Motion Model
Ekkasit Pinyoanuntapong (University of North Carolina at Charlotte), Chen Chen (University of Central Florida)
GenerationData SynthesisTransformerVision-Language-Action ModelAuto EncoderTextMultimodalitySequential
🎯 What it does: This paper proposes a Bidirectional Autoregressive Motion Model (BAMM), enabling the generation and editing of text-to-3D human motion without requiring prior knowledge of motion length.
BaSIC: BayesNet Structure Learning for Computational Scalable Neural Image Compression
Yufeng Zhang (Shanghai Jiao Tong University), Weiyao Lin (Shanghai Jiao Tong University)
CompressionImage
🎯 What it does: Propose a computable and scalable neural image compression framework called BaSIC based on Bayesian network structure learning;
Bayesian Detector Combination for Object Detection with Crowdsourced Annotations
Zhi Qin Tan (University of Surrey), Yunpeng Li (University of Surrey)
Object DetectionImageBiomedical Data
🎯 What it does: This paper proposes a framework named Bayesian Detector Combination (BDC) for training object detection models on noisy data with multiple annotators, and can automatically infer the quality of each annotator.
Bayesian Evidential Deep Learning for Online Action Detection
Hongji Guo (Rensselaer Polytechnic Institute), Qiang Ji (Rensselaer Polytechnic Institute)
RecognitionKnowledge DistillationTransformerVideo
🎯 What it does: This paper proposes a Bayesian Evidential Deep Learning (BEDL) framework for online action detection, transferring posterior distribution knowledge from a teacher-student structure to an evidence learning model to achieve efficient prediction and uncertainty quantification.
Bayesian Self-Training for Semi-Supervised 3D Segmentation
Ozan Unal (ETH Zurich), Luc Van Gool (ETH Zurich)
SegmentationPoint CloudMesh
🎯 What it does: This paper proposes a Bayesian self-training based semi-supervised 3D segmentation framework, applicable to 3D semantic segmentation, instance segmentation, and dense 3D visual localization;
Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation
Omer Dahary (Tel Aviv University), Danny Cohen-Or (Tel Aviv University)
GenerationDiffusion modelImageTextMultimodality
🎯 What it does: Propose a training-free "Bounded Attention" method, which suppresses semantic leakage during multi-agent text-to-image generation by introducing time-specific masks in cross-attention and self-attention layers, thereby achieving more precise layout control.
Be-Your-Outpainter: Mastering Video Outpainting through Input-Specific Adaptation
Fu-Yun Wang (CUHK), Hongsheng Li (CUHK)
GenerationTransformerDiffusion modelVideo
🎯 What it does: Proposes MOTIA, an open-domain video outpainting framework based on diffusion models, capable of handling videos with arbitrary masks, resolutions, and styles. The framework achieves this through a two-stage process: ① Input-specific adaptation stage, where pseudo-outpainting is learned on the source video using a lightweight LoRA adapter to capture video-specific motion and content patterns; ② Mode-aware outpainting stage, which combines the learned patterns with the prior knowledge of diffusion models, employing spatially aware insertion (SA-Insertion) and noise regret techniques to enhance edge filling quality and spatiotemporal consistency.
BEAF: Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models
Moon Ye-Bin (POSTECH), Tae-Hyun Oh (POSTECH)
Large Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: Proposed the BEAF (BEfore-AFter) benchmark, which evaluates hallucination behaviors in Vision-Language Models (VLMs) by leveraging before-and-after changes in visual scenes.
Beat-It: Beat-Synchronized Multi-Condition 3D Dance Generation
Zikai Huang (South China University of Technology), Shengfeng He (Singapore Management University)
GenerationData SynthesisTransformerDiffusion modelVideoMultimodalityAudio
🎯 What it does: Propose the Beat-It framework, achieving multi-condition 3D dance generation with three conditions: music, key poses, and beats, while improving beat synchronization and pose control through explicit beat representation and hierarchical fusion.
BenchLMM: Benchmarking Cross-style Visual Capability of Large Multimodal Models
Rizhao Cai (Nanyang Technological University), Alex Kot (Nanyang Technological University)
Prompt EngineeringVision Language ModelImageBiomedical DataMagnetic Resonance ImagingComputed TomographyBenchmark
🎯 What it does: Propose BenchLMM benchmark for quantitatively assessing the reasoning robustness of large multimodal models across three visual styles (art, sensor, and application), and systematically evaluate existing open-source models and GPT-4V.
Benchmarking Object Detectors with COCO: A New Path Forward
Shweta Singh (IIT Roorkee), Karan Desai (University of Michigan)
Object DetectionImageBenchmark
🎯 What it does: Provide higher-quality annotations for COCO-2017 masks and construct the COCO-ReM dataset.
Benchmarking Spurious Bias in Few-Shot Image Classifiers
Guangtao Zheng (University of Virginia), Aidong Zhang (University of Virginia)
ClassificationVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed a framework named FewSTAB, an automated benchmark framework for systematically evaluating the robustness of few-shot image classifiers against spurious bias (dirty attributes).
Benchmarking the Robustness of Cross-view Geo-localization Models
Qingwang Zhang (Shenzhen University), Yingying Zhu (Shenzhen University)
Data SynthesisRetrievalImageBenchmark
🎯 What it does: Establish a robustness benchmark for cross-view geolocation models by generating approximately 1.5 million corrupted images and evaluating model performance.
Benchmarks and Challenges in Pose Estimation for Egocentric Hand Interactions with Objects
Zicong Fan (ETH Zürich), Angela Yao (National University of Singapore)
Pose EstimationTransformerImageBenchmark
🎯 What it does: This paper proposes and organizes the HANDS23 public challenge, based on the AssemblyHands and ARCTIC datasets, which provide egocentric views of hand-object interactions. It offers a unified training/test split and establishes evaluation standards for two tasks: 3D hand pose estimation (single-view) and hand-object consistent motion reconstruction (allocentric/egocentric). The paper collects and systematically evaluates participating methods, analyzing the performance improvements from techniques such as multi-view fusion, camera distortion compensation, and Transformer visual models, while highlighting existing challenges in scenarios involving rapid motion, narrow perspectives, and close hand-object contact.
BeNeRF:Neural Radiance Fields from a Single Blurry Image and Event Stream
Wenpu Li (West Lake University), Peidong Liu (West Lake University)
RestorationGenerationNeural Radiance FieldImage
🎯 What it does: This paper proposes a method called BeNeRF, which can recover neural radiance fields (NeRF) and the continuous motion trajectory of the camera from a single blurry image and its corresponding event stream, thereby synthesizing high-quality clear images and videos.
Beta-Tuned Timestep Diffusion Model
Tianyi Zheng (Shanghai Jiao Tong University), Bo Li (vivo Mobile Communication Co., Ltd)
GenerationDiffusion modelImage
🎯 What it does: Proposes the Beta-Tuned Timestep Diffusion Model (B-TTDM), which uses the Beta distribution to perform non-uniform adjustment of time step sampling during diffusion model training, better matching the distribution changes in the forward diffusion process.
Betrayed by Attention: A Simple yet Effective Approach for Self-supervised Video Object Segmentation
Shuangrui Ding (Chinese University of Hong Kong), Hongkai Xiong (Shanghai Jiao Tong University)
SegmentationTransformerContrastive LearningVideo
🎯 What it does: This paper proposes a fully self-supervised video object segmentation (VOS) method BA, which leverages the attention maps from DINO pre-trained ViT, learns spatiotemporal correspondence through a single spatiotemporal Transformer block, and directly obtains object segmentation masks in videos using hierarchical clustering.
Better Call SAL: Towards Learning to Segment Anything in Lidar
Aljosa Osep, Laura Leal-Taixé
SegmentationConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodalityPoint Cloud
🎯 What it does: Propose the SAL (Segment Anything in Lidar) model, achieving zero-shot point cloud instance segmentation and classification based on text prompts;
Better Regression Makes Better Test-time Adaptive 3D Object Detection
Jiakang Yuan (Fudan University), Tao Chen (Fudan University)
Object DetectionDomain AdaptationAutonomous DrivingPoint CloudBenchmark
🎯 What it does: Proposed a test-time domain adaptation method called Reg-TTA3D for 3D object detection, specifically addressing cross-domain challenges in dynamic environments.
Beyond MOT: Semantic Multi-Object Tracking
Yunhao Li (Institute of Software Chinese Academy of Sciences), Libo Zhang (Institute of Software Chinese Academy of Sciences)
RecognitionObject TrackingGenerationConvolutional Neural NetworkTransformerVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes the semantic multi-object tracking (SMOT) task, which involves tracking target trajectories while simultaneously providing instance description, instance interaction recognition, and video-level description for each trajectory; based on this, a large-scale semantic multi-object tracking benchmark called BenSMOT is constructed, offering four types of annotations: trajectories, instance descriptions, interaction labels, and video descriptions; finally, an end-to-end trained baseline model named SMOTer is presented.
Beyond Pixels: Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier
Prantik Howlader (Stony Brook University), Dimitris Samaras (Stony Brook University)
SegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a multi-scale patch multi-label classifier (MPMC) as a plug-in module to enhance contextual information in semi-supervised semantic segmentation models, and reduce noise generated by the teacher model through confidence learning to adaptively adjust pseudo-label weights.
Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free Continual Learning
Xinyuan Gao (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringImageBenchmark
🎯 What it does: Propose an example-free continual learning framework based on Continuous Adapter (C-ADA), replacing traditional prompt learning and key-query matching mechanisms.
Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models
Hyeonwoo Kim (Seoul National University), Hanbyul Joo (Naver Webtoon AI)
GenerationData SynthesisPose EstimationDiffusion modelImagePoint CloudMesh
🎯 What it does: This paper proposes a new comprehensive affordance (ComA) representation that can simultaneously model contact, relative pose, and spatial relationships in human-object interaction.
Beyond the Data Imbalance: Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction
Hyeongseok Jeon (MORAI Inc.), Dongsuk Kum (Korea Advanced Institute of Science and Technology)
Autonomous DrivingRecurrent Neural NetworkGraph Neural NetworkContrastive LearningVideoSequential
🎯 What it does: This paper proposes a two-stage vehicle maneuver prediction framework. It first learns maneuver representations using a balanced intersection viewpoint dataset, then learns geometric diversity using a self-vehicle perspective dataset. Finally, it calculates maneuver probabilities during inference using Gaussian kernel density estimation.
Beyond Viewpoint: Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part Representation
Linlong Fan (University of Electronic Science and Technology of China), Lixin Duan (University of Electronic Science and Technology of China)
RecognitionConvolutional Neural NetworkTransformerImagePoint CloudBenchmark
🎯 What it does: Propose Part-Aware Network (PANet), which leverages weakly supervised part localization and adaptive part refinement to extract global part representations from multi-view images, achieving robust 3D object recognition under arbitrary viewpoints.
BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion
Gwanghyun Kim (Seoul National University), Se Young Chun (Seoul National University)
GenerationDiffusion modelImageText
🎯 What it does: Proposed the BeyondScene framework, achieving human-centric scene generation with ultra 8K resolution through a two-stage hierarchical generation method based on pre-trained diffusion models, overcoming challenges such as text description token limitations, training image size limitations, and multi-human scene generation.
Bi-directional Contextual Attention for 3D Dense Captioning
Minjung Kim (Seoul National University), Gunhee Kim (Seoul National University)
Object DetectionTransformerLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: This paper proposes the BiCA model, which utilizes bidirectional context attention to achieve object localization in 3D scenes and natural language descriptions for each object.
BI-MDRG: Bridging Image History in Multimodal Dialogue Response Generation
Hee Suk Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: In the multi-modal dialogue response generation task, this paper proposes the BI-MDRG model, which can simultaneously generate text and image responses, and enhances the image-rootedness of text responses and the consistency of image responses through bridging image history.
Bi-TTA: Bidirectional Test-Time Adapter for Remote Physiological Measurement
Haodong LI, Yingcong Chen (Hong Kong University Of Science And Technology)
Domain AdaptationConvolutional Neural NetworkVideoBenchmark
🎯 What it does: Propose a bidirectional test-time adaptation framework called Bi-TTA that performs single-instance adaptation using only user facial video during inference, addressing the generalization problem of remote photoplethysmography (rPPG) models in unseen domains.
Bidirectional Progressive Transformer for Interaction Intention Anticipation
Zichen Zhang (University of Science and Technology of China), Yang Cao (University of Science and Technology of China)
Object TrackingPose EstimationTransformerAuto EncoderVideoBenchmark
🎯 What it does: Propose a bidirectional progressive Transformer (BOT) that simultaneously predicts future hand trajectories and interaction hotspots in first-person videos, enhancing prediction accuracy through a bidirectional mutual correction mechanism.
Bidirectional Stereo Image Compression with Cross-Dimensional Entropy Model
Zhening Liu (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
CompressionConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented a fully symmetric bidirectional stereo image compression framework called BiSIC, which utilizes 3D convolutions to extract local and disparity information, incorporates bidirectional mutual attention blocks to share global features, and proposes a cross-dimensional entropy model that integrates hyperprior, spatial context, channel context, and stereo dependencies, ultimately achieving efficient compression.
Bidirectional Uncertainty-Based Active Learning for Open-Set Annotation
Chen-Chen Zong (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
ClassificationAnomaly DetectionOptimizationData-Centric LearningImageBenchmark
🎯 What it does: In open-set annotation tasks, a bidirectional uncertainty active learning framework (BUAL) was designed, which pushes unknown class samples toward high confidence intervals through random label negative learning, and selects the most valuable known class samples by combining bidirectional uncertainty sampling from positive and negative classification heads.
Binomial Self-compensation for Motion Error in Dynamic 3D Scanning
Geyou Zhang (University of Electronic Science and Technology of China), Kai Liu (Sichuan University)
Depth EstimationImageVideoPoint Cloud
🎯 What it does: To address motion errors in dynamic 3D scanning, this paper proposes a binomial self-compensation (BSC) algorithm based on four-step phase-shifting phase measurement (PSP), achieving cyclic projection under a set of high-frequency phase-shifting patterns and real-time point cloud reconstruction.
BKDSNN: Enhancing the Performance of Learning-based Spiking Neural Networks Training with Blurred Knowledge Distillation
Zekai Xu (Shanghai Jiao Tong University), Zhezhi He (Shanghai Jiao Tong University)
ClassificationKnowledge DistillationConvolutional Neural NetworkSpiking Neural NetworkImageSequential
🎯 What it does: Propose a fuzzy knowledge distillation (BKD) technique, mapping randomly blurred SNN features through a recovery block to continuous features, thereby more effectively distilling features from the teacher ANN, and combining with logits distillation to enhance the performance of learning-based spiking neural networks (SNNs).
BlazeBVD: Make Scale-Time Equalization Great Again for Blind Video Deflickering
Xinmin Qiu (University of Chinese Academy of Sciences), Xuecheng Nie (Sichuan University)
RestorationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: Propose a histogram-assisted blind video deflickering (BVD) method called BlazeBVD, which can rapidly remove global and local flickering and enhance video temporal consistency without prior knowledge of flicker types.
BlenderAlchemy: Editing 3D Graphics with Vision-Language Models
Ian Huang (Stanford University), Leonidas Guibas (Stanford University)
GenerationData SynthesisLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMesh
🎯 What it does: Propose the BlenderAlchemy system, which leverages visual language models (e.g., GPT-4V) and image generation models (e.g., DALL-E3) to perform multi-candidate generation and visual evaluation of Blender visual programs, iteratively optimizing the programs to achieve material, geometry, and lighting edits based on text or reference images.
Blind image deblurring with noise-robust kernel estimation
Chanseok Lee (KAIST), Mooseok Jang (KAIST)
RestorationConvolutional Neural NetworkImage
🎯 What it does: Studied the problem of blind image deblurring in strong noise environments, and proposed an end-to-end blind deblurring framework based on a noise-robust convolution kernel estimation function and deep image prior (DIP).
Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding
Jiangtao Zhang (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
RestorationGenerative Adversarial NetworkImageBenchmark
🎯 What it does: Propose a blind image deconvolution framework that utilizes generative adversarial networks (GANs) to learn the distribution of blur kernels and obtain kernel initialization through latent space mapping, followed by jointly optimizing the image and kernel via a deep network within the latent kernel space.
BLINK: Multimodal Large Language Models Can See but Not Perceive
Xingyu Fu (University of Pennsylvania), Ranjay Krishna (University of Washington)
ClassificationRecognitionRestorationPose EstimationDepth EstimationLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Create the Blink visual perception benchmark, which includes 14 classical computer vision tasks converted into 3,807 multiple-choice questions, and evaluates the performance of 16 multimodal large language models.
Bones Can't Be Triangles: Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision
Jinhee Kim (Korea Advanced Institute of Science and Technology), Jaegul Choo (Korea Advanced Institute of Science and Technology)
Data SynthesisPose EstimationConvolutional Neural NetworkBiomedical Data
🎯 What it does: By constructing the KeyBot framework, automatically detect and correct typical errors in spinal key points on X-ray images, thereby reducing human intervention before interactive key point estimation.
Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering
Francesco Di Sario (University of Turin), Enzo Tartaglione (LTCI Télécom Paris Institut Polytechnique de Paris)
GenerationComputational EfficiencyMixture of ExpertsNeural Radiance FieldImage
🎯 What it does: This paper proposes a model-agnostic sparse Mixture-of-Experts (MoE) framework that dynamically routes sampling points to different resolution Fast-NeRF models to enhance rendering quality without significantly increasing computational cost.
Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training
Qiangqiang Wu (City University of Hong Kong), Antoni Chan
Object TrackingAutonomous DrivingKnowledge DistillationTransformerImagePoint Cloud
🎯 What it does: Propose a unified 3D single-target tracking framework that enhances performance through 3D pre-training and 2D matching distillation
Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model
Yang Jin (Xi'an University of Architecture and Technology), Binglu Wang (Xi'an University of Architecture and Technology)
Object DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: This paper proposes an end-to-end Gaze Object Segmentation (GOS) framework that generates pixel-level masks using a visual foundation model (SAM), and accurately predicts pixel-level masks of human-gazed objects through a unified detection and segmentation Transformer, head feature reconstruction, and a spatial-to-target gaze regression strategy.
Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training
Cheng Tan (Zhejiang University), Stan Z. Li (Westlake University)
TransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a multimodal reasoning method based on self-consistency training called MC-CoT. During training, multiple reasoning chains and answers are generated through multiple sampling, and a voting mechanism is used to select the most consistent and reliable reasoning chains and answers, significantly improving the model's reasoning quality.
Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory
Sensen Gao (Nankai University), Qing Guo (Agency for Science, Technology and Research)
Adversarial AttackVision Language ModelMultimodality
🎯 What it does: Propose a novel method to enhance the transferability of adversarial attacks in vision-language pre-training models: by sampling from the cross-region of adversarial trajectories and combining text-guided sample selection to generate diverse and more transferable multimodal adversarial examples; simultaneously, introduce a cross-region offset strategy in the text modality to further reduce the risk of overfitting on the target model.
Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing
Siqi Liu, Pong C. Yuen (Hong Kong Baptist University)
Domain AdaptationAnomaly DetectionSafty and PrivacyTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: Enhance the generalization capability of CLIP in facial anti-spoofing through low-level domain prompt tuning
Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals
Camilo L Fosco, Aude Oliva (Massachusetts Institute of Technology)
GenerationTransformerDiffusion modelContrastive LearningVideoTextBiomedical DataMagnetic Resonance ImagingRetrieval-Augmented Generation
🎯 What it does: Developed a three-stage pipeline to map fMRI signals to the latent space of a video generation model and text conditions, successfully reconstructing 2-3 second short videos.
Brain-ID: Learning Contrast-agnostic Anatomical Representations for Brain Imaging
Peirong Liu (Harvard Medical School and Massachusetts General Hospital), Juan E. Iglesias (Harvard Medical School and Massachusetts General Hospital)
RestorationSegmentationData SynthesisSuper ResolutionConvolutional Neural NetworkBiomedical DataMagnetic Resonance ImagingComputed TomographyAlzheimer's Disease
🎯 What it does: This paper proposes Brain-ID, a contrast-agnostic anatomical feature learning framework based on synthetic data generation, for unified representation in brain imaging.
BRAVE: Broadening the visual encoding of vision-language models
Oğuzhan Fatih Kar (Google), Federico Tombari (Google)
Representation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose the BRAVE framework, integrating features from multiple visual encoders to generate compressed visual representations as soft prompts input into a frozen language model, thereby significantly enhancing the visual capabilities of vision-language models.
Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection
Qijie Mo (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the information asymmetry problem in incremental object detection, proposing the Bridge Past and Future (BPF) method, which combines pseudo labels from the old model, uses feature attention to exclude future class backgrounds, and designs a Distillation with Future (DwF) loss to preserve old class knowledge and learn new classes.
BRIDGE: Bridging Gaps in Image Captioning Evaluation with Stronger Visual Cues
Sara Sarto (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
RecognitionTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed a new, reference-free image caption evaluation metric called BRIDGE, which integrates fine-grained visual features into multimodal pseudo-descriptions to more accurately measure the correspondence between generated text and images.
Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation
Shihao Zhao (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)
GenerationTransformerSupervised Fine-TuningDiffusion modelMultimodality
🎯 What it does: Propose the LaVi-Bridge framework, achieving seamless integration of any pre-trained language model with a generative vision model for text-to-image diffusion generation tasks.
Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors
Tongkun Guan (Sun Yat-sen University), Xiaokang Yang (Sun Yat-sen University)
RestorationGenerative Adversarial NetworkImage
🎯 What it does: Proposes a framework for image super-resolution that utilizes user multi-scale inputs and learnable masks to significantly enhance image quality while preserving details.
Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases
Xinpeng Liu (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
GenerationData SynthesisPose EstimationTransformerDiffusion modelAuto EncoderTextSequentialBenchmark
🎯 What it does: This paper proposes an intermediate representation between motion and action semantics—Kinematic Phrases (KP), and constructs a unified large-scale motion knowledge base, motion understanding system, and white-box evaluation benchmark for motion generation.
Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture
ShahRukh Athar (Captions Research), Chen Cao (Meta Reality Labs)
Image TranslationRestorationGenerationDiffusion modelGenerative Adversarial NetworkImageVideo
🎯 What it does: Extract facial texture maps from monocular short videos captured by smartphones, reconstruct lighting and missing regions in the W+ space using StyleGAN2, then enhance details with a conditional diffusion model based on texture gradients, generating high-resolution, uniformly illuminated facial textures, and subsequently producing photorealistic animated avatars.
Bridging the Pathology Domain Gap: Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data
Zhengfeng Lai (University of California, Davis), Chen-Nee Chuah (University of California, Davis)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Proposes a framework named Path-CLIP for efficiently fine-tuning the CLIP model in pathological image analysis tasks with only a minimal amount of labeled data.
BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion
Xuan Ju (Tencent PCG), Qiang Xu (Chinese University of Hong Kong)
RestorationConvolutional Neural NetworkDiffusion modelAuto EncoderImageMultimodalityBenchmark
🎯 What it does: Propose a dual-branch, plug-and-play image inpainting model called BrushNet, which can hierarchically inject pixel-level occlusion information into pre-trained diffusion models, achieving high-quality, semantically consistent image restoration.
Bucketed Ranking-based Losses for Efficient Training of Object Detectors
Feyza Yavuz (METU), Sinan Kalkan (METU)
Object DetectionComputational EfficiencyTransformerContrastive LearningImage
🎯 What it does: This paper proposes Bucketed Ranking-based Losses (BRS), which significantly reduces the time and space complexity of ranking-based loss functions (such as AP Loss, RS Loss) by dividing negative samples into fewer buckets after sorting them by scores.
BugNIST - a Large Volumetric Dataset for Detection under Domain Shift
Patrick M Jensen, Anders B Dahl
Object DetectionData SynthesisConvolutional Neural NetworkBiomedical DataComputed TomographyBenchmark
🎯 What it does: Designed and released the BugNIST dataset as a benchmark for voxel object detection under context shift (domain shift), and evaluated the baseline performance of three existing 3D detection models on this dataset.
BurstM: Deep Burst Multi-scale SR using Fourier Space with Optical Flow
EungGu Kang (Daegu Gyeongbuk Institute of Science and Technology), Kyong Hwan Jin (Korea University)
Super ResolutionOptical FlowImage
🎯 What it does: Proposed a multi-frame super-resolution network named BurstM, which can precisely align low-resolution burst images without using DCN by leveraging optical flow, and reconstructs high-frequency textures by predicting continuous Fourier coefficients in the Fourier space; simultaneously supports multi-scale (×2, ×3, ×4) super-resolution under a single model;
C2C: Component-to-Composition Learning for Zero-Shot Compositional Action Recognition
Rongchang Li (Jiangnan University), Josef Kittler (University of Surrey)
RecognitionTransformerVision-Language-Action ModelContrastive LearningVideoBenchmark
🎯 What it does: Propose the Zero-Shot Compositional Action Recognition (ZS-CAR) task, construct the corresponding Sth-com benchmark dataset, and develop the Component-to-Composition (C2C) learning framework to accomplish this task.
CadVLM: Bridging Language and Vision in the Generation of Parametric CAD Sketches
Sifan Wu (Université de Montréal & MILA), Bang Liu (Autodesk AI Lab)
GenerationTransformerVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a multimodal vision-language model called CadVLM for generating parameterized CAD sketches, enabling tasks such as automatic completion, automatic constraint generation, and image-conditioned generation.
CaesarNeRF: Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering
Haidong Zhu (University of Southern California), Luming Liang (Microsoft)
GenerationRepresentation LearningConvolutional Neural NetworkTransformerNeural Radiance FieldImage
🎯 What it does: Propose CaesarNeRF, a generalizable NeRF framework capable of achieving multi-view rendering using only a single reference image;
Caltech Aerial RGB-Thermal Dataset in the Wild
Connor Lee (California Institute of Technology), Soon-Jo Chung (California Institute of Technology)
Image TranslationSegmentationRobotic IntelligenceImageMultimodalityBenchmark
🎯 What it does: Proposed the first publicly available RGB-thermal imaging dataset aimed at supporting aerial robots operating in natural environments. The dataset contains synchronized RGB, thermal imaging, global positioning, and inertial data, along with semantic segmentation annotations for 10 common categories.
Camera Calibration using a Collimator System
Shunkun Liang (National University of Defense Technology), Yang Shang (National University of Defense Technology)
OptimizationImagePhysics Related
🎯 What it does: Proposed a camera calibration method based on a collimator system, leveraging the angle invariance of the collimator to constrain the relative motion between the camera and the calibration board to spherical motion, thereby enabling estimation of camera intrinsic parameters with only a few poses;
Camera Height Doesn't Change: Unsupervised Training for Metric Monocular Road-Scene Depth Estimation
Genki Kinoshita (Kyoto University), Ko Nishino (Kyoto University)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: Proposes the FUMET framework, which enables monocular depth networks to learn absolute scale and achieve metric road scene depth estimation by leveraging vehicle dimensions on roads as unsupervised scale supervision.
Camera-LiDAR Cross-modality Gait Recognition
Wenxuan Guo (Tsinghua University), Jie Zhou (Tsinghua University)
RecognitionDepth EstimationConvolutional Neural NetworkContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: Propose the first camera-radar cross-modal gait recognition framework CL-Gait, which extracts 2D gait silhouettes and 3D point cloud features using dual-stream networks, and achieves cross-modal matching through alignment learning.
CamoTeacher: Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection
xunfa lai, Rongrong Ji (Contemporary Amperex Technology Co. Limited)
Object DetectionImageBenchmark
🎯 What it does: Propose the CamoTeacher framework, which combines semi-supervised learning with pseudo-label consistency learning for camouflaged object detection; improve the training effectiveness of the student model by dynamically balancing pixel-level and instance-level noise in pseudo-labels through Dual-Rotation Consistency Learning (DRCL).
Can OOD Object Detectors Learn from Foundation Models?
Jiahui Liu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
Object DetectionData SynthesisAnomaly DetectionConvolutional Neural NetworkLarge Language ModelDiffusion modelImage
🎯 What it does: This paper leverages large language models, text-to-image diffusion models, and segmentation models to automatically generate and annotate scene-level synthetic OOD samples, aiming to enhance OOD object detection performance.
Can Textual Semantics Mitigate Sounding Object Segmentation Preference?
Yaoting Wang (Renmin University of China), Di Hu (Renmin University of China)
SegmentationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityAudio
🎯 What it does: Proposes the TeSO method, which leverages text semantics to enhance audio-visual segmentation through modules such as image description, frozen LLM inference of potential sound objects, semantic-driven audio modeling (SeDAM), and masked query prompting (PMQS).
Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning
Ali Cheraghian (Data61, CSIRO), Mehrtash Harandi (Monash University)
ClassificationTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextPoint Cloud
🎯 What it does: This paper proposes a complete framework called C3PR for few-shot incremental learning (FSCIL) of 3D point clouds using the CLIP model.
CanonicalFusion: Generating Drivable 3D Human Avatars from Multiple Images
Jisu Shin (GIST AI Graduate School), Hae-Gon Jeon (GIST AI Graduate School)
GenerationDepth EstimationConvolutional Neural NetworkAuto EncoderImageMesh
🎯 What it does: Propose CanonicalFusion, which simultaneously predicts depth and compressed linear blend skinning (LBS) weights using multiple images, and generates a drivable 3D human model based on this;
CARB-Net: Camera-Assisted Radar-Based Network for Vulnerable Road User Detection
Wei-Yu Lee (Ghent University), Wilfried Philips (Ghent University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkImageMultimodalityPoint Cloud
🎯 What it does: Proposed CARB-Net, which integrates camera angular precision with radar depth perception, achieving breakthroughs in addressing the challenge of radar's inability to distinguish nearby targets when angular resolution is insufficient.
CardiacNet: Learning to Reconstruct Abnormalities for Cardiac Disease Assessment from Echocardiogram Videos
Jiewen Yang (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
RestorationAnomaly DetectionConvolutional Neural NetworkAuto EncoderGenerative Adversarial NetworkVideoBiomedical DataUltrasound
🎯 What it does: Propose CardiacNet, a reconstruction-based deep learning framework for learning local structural and motion abnormalities from cardiac ultrasound videos and assessing cardiac diseases.
CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting
Jiezhi Yang, Joseph E Gonzalez
Autonomous DrivingConvolutional Neural NetworkTransformerNeural Radiance FieldAuto EncoderImage
🎯 What it does: Proposed a method called CARFF for predicting future 3D scenes in partially observable environments, supporting autonomous driving decision-making through predicted 3D views.
CarFormer: Self-Driving with Learned Object-Centric Representations
Shadi Hamdan (Koç University), Fatma Guney
Autonomous DrivingRecurrent Neural NetworkTransformerAuto EncoderWorld ModelVideoSequentialBenchmark
🎯 What it does: This paper proposes CarFormer, an autoregressive transformer model that extracts object information from bird's-eye view (BEV) sequences using self-supervised learning-based object-centric slot representations, and learns driving behaviors and predicts scene dynamics based on this.
Cascade Prompt Learning for Visual-Language Model Adaptation
Ge Wu (Nankai University), Xiang Li (Nankai University)
ClassificationDomain AdaptationKnowledge DistillationPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Propose a two-stage Prompt learning framework called CasPL, which enhances domain-agnostic knowledge and adapts to task-specific knowledge on vision-language models (e.g., CLIP), and seamlessly integrates as a plugin into existing Prompt methods.
Cascade-Zero123: One Image to Highly Consistent 3D with Self-Prompted Nearby Views
Yabo Chen (Shanghai Jiao Tong University), Qi Tian (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelImageMesh
🎯 What it does: This paper proposes the Cascade-Zero123 method, which generates high-consistency 3D views from a single image using self-prompted nearby views in a cascading manner.
CAT-SAM: Conditional Tuning for Few-Shot Adaptation of Segment Anything Model
Aoran Xiao (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
SegmentationComputational EfficiencyMeta LearningTransformerSupervised Fine-TuningPrompt EngineeringImageBiomedical Data
🎯 What it does: Efficiently adapt the Segment Anything Model (SAM) under conditions of using only a minimal number of labeled samples (single-sample or a few samples) to achieve various downstream segmentation tasks.
CAT: Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios
Qilang Ye, Xiaochun Cao
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: This paper designs and trains the CAT (Clue-Aggregated Transformer) multimodal large language model for answering questions in dynamic audio-visual scenarios, improving answer quality through question clue aggregation and ambiguity elimination.
Catastrophic Overfitting: A Potential Blessing in Disguise
MN Zhao (Dalian University of Technology), Baocai Yin (Dalian University of Technology)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper conducts an in-depth analysis of the catastrophic overfitting (CO) phenomenon observed in Fast Adversarial Training (FAT), and proposes a regularization method based on feature activation differences to suppress or induce CO. Subsequently, it employs CO models during inference by adding random noise to achieve 'attack fuzzification,' thereby improving robustness against adversarial samples while maintaining high accuracy on clean samples.
CatchBackdoor: Backdoor Detection via Critical Trojan Neural Path Fuzzing
Haibo Jin (University of Illinois Urbana-Champaign), Haohan Wang (University of Illinois Urbana-Champaign)
Anomaly DetectionSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Propose the CatchBackdoor method, which utilizes neural path differential fuzzing to detect backdoors in deep neural networks, capable of reversing triggers and inducing erroneous predictions without trigger samples;
Category Adaptation Meets Projected Distillation in Generalized Continual Category Discovery
Grzegorz Rypeść (IDEAS NCBR), Bartlomiej Twardowski
Knowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes a method called CAMP, which integrates category adaptation with projection distillation for generalized continual category discovery (GCCD) and sample-free memory category incremental learning.
Causal Subgraphs and Information Bottlenecks: Redefining OOD Robustness in Graph Neural Networks
Weizhi An (University of Texas at Arlington), Junzhou Huang (University of Texas at Arlington)
Domain AdaptationRepresentation LearningGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes an end-to-end CSIB framework that leverages causal structure models and Graph Information Bottleneck (GIB) to extract environment-invariant causal subgraphs from graphs, thereby enhancing robustness to out-of-distribution (OOD) scenarios.
Causality-inspired Discriminative Feature Learning in Triple Domains for Gait Recognition
Haijun Xiong (Huazhong University of Science and Technology), Wenyu Liu (Huazhong University of Science and Technology)
RecognitionRepresentation LearningContrastive LearningVideo
🎯 What it does: This paper proposes the CLTD (Causality-inspired Discriminative Feature Learning in Triple Domains) module, which utilizes causal intervention and contrastive learning to remove non-identity factors in the spatial, temporal, and spectral domains, thereby enhancing the discriminative performance of gait recognition.
CC-SAM: Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation
Shreyank N Gowda (University of Oxford), David A Clifton
SegmentationConvolutional Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataUltrasound
🎯 What it does: Proposed CC-SAM, an improved base model for ultrasound image segmentation, enhancing SAM's performance in medical imaging through freezing the CNN branch with ViT fusion, variational attention fusion, and text prompts generated by ChatGPT.
cDP-MIL: Robust Multiple Instance Learning via Cascaded Dirichlet Process
Yihang Chen (University of Hong Kong), Lequan Yu (University of Hong Kong)
ClassificationAnomaly DetectionContrastive LearningImageBiomedical DataBenchmark
🎯 What it does: Proposed a multi-instance learning framework based on cascading Dirichlet process (cDP-MIL), which clusters WSI local patches using deep variational parameter DP, generates sliding window-level representations, and provides uncertainty estimation and tumor localization;
Centering the Value of Every Modality: Towards Efficient and Resilient Modality-agnostic Semantic Segmentation
Xu Zheng (Hong Kong University of Science and Technology), Lin Wang (Hong Kong University of Science and Technology)
SegmentationComputational EfficiencyTransformerMultimodality
🎯 What it does: Proposes a multi-modal semantic segmentation framework named MAGIC, which achieves efficient, robust, and modality-agnostic semantic segmentation by utilizing a Multi-Modal Aggregation Module (MAM) and an Arbitrary Modal Selection Module (ASM).