ICCV 2025 Papers — Page 5
IEEE/CVF International Conference on Computer Vision · 2701 papers
CompSlider: Compositional Slider for Disentangled Multiple-Attribute Image Generation
Zixin Zhu (University at Buffalo), Junsong Yuan (University at Buffalo)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: This paper presents CompSlider, a combined slider model that enables discrete and continuous control of multiple attributes (such as age, expression, posture, style, etc.) in text-to-image diffusion models, allowing for the adjustment of multiple attributes simultaneously while maintaining image structural consistency in a single forward pass.
ConceptSplit: Decoupled Multi-Concept Personalization of Diffusion Models via Token-wise Adaptation and Attention Disentanglement
Habin Lim (Korea University), Gyeong-Moon Park (Korea University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes the ConceptSplit framework, achieving non-fusion training and inference of multi-concept personalized diffusion models.
Conditional Latent Diffusion Models for Zero-Shot Instance Segmentation
Maximilian Ulmer (German Aerospace Center), Maximilian Durner (Technical University of Munich)
Object DetectionSegmentationTransformerDiffusion modelImage
🎯 What it does: This paper proposes the Object Conditional Diffusion Transformer (OC-DiT), a zero-shot instance segmentation method based on diffusion models that can generate instance masks without the need for training on target objects.
Conditional Visual Autoregressive Modeling for Pathological Image Restoration
Ziyi Liu (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
RestorationSuper ResolutionTransformerContrastive LearningImage
🎯 What it does: This paper studies a conditional visual autoregressive model called CVARPath for denoising, deblurring, and super-resolution of pathological images, which can generate high-quality images from degraded ones.
ConformalSAM: Unlocking the Potential of Foundational Segmentation Models in Semi-Supervised Semantic Segmentation with Conformal Prediction
Danhui Chen (Dalian University of Technology), Xiangyang Ji (Tsinghua University)
SegmentationSupervised Fine-TuningImage
🎯 What it does: Using the basic segmentation model SEEM to generate masks for unlabeled images, and performing self-training in semi-supervised semantic segmentation tasks after calibration through conformal prediction.
Confound from All Sides, Distill with Resilience: Multi-Objective Adversarial Paths to Zero-Shot Robustness
Junhao Dong (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
OptimizationKnowledge DistillationAdversarial AttackSupervised Fine-TuningVision Language ModelImageMultimodalityBiomedical Data
🎯 What it does: This study investigates the generation of diverse adversarial samples through multi-objective optimization to achieve robust knowledge distillation from large visual language models (VLM) to lightweight models, enhancing zero-shot robustness.
Consensus-Driven Active Model Selection
Justin Kay (Massachusetts Institute of Technology), Sara Beery (Massachusetts Institute of Technology)
ClassificationBenchmark
🎯 What it does: A consensus-driven active model selection method called CODA is proposed, which quickly identifies the best pre-trained model using a small number of labels.
Consistency Trajectory Matching for One-Step Generative Super-Resolution
Weiyi You (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
RestorationGenerationSuper ResolutionDiffusion modelGenerative Adversarial NetworkImageOrdinary Differential Equation
🎯 What it does: A model for generating super-resolution images in a single step without distillation (CTMSR) is proposed, which learns a deterministic mapping from low-resolution images with noise to high-resolution images through consistency training, and further enhances the naturalness of generated images using Distribution Trajectory Matching (DTM).
Consistent Time-of-Flight Depth Denoising via Graph-Informed Geometric Attention
Weida Wang (Tongji University), Di Qiu (Google)
RestorationDepth EstimationGraph Neural NetworkVideo
🎯 What it does: This paper proposes a time-of-flight (ToF) depth denoising network based on cross-frame graph structure fusion—GIGA-ToF, which enhances temporal consistency while preserving spatial details.
ConsistentCity: Semantic Flow-guided Occupancy DiT for Temporally Consistent Driving Scene Synthesis
Benjin Zhu (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationData SynthesisAutonomous DrivingTransformerDiffusion modelImageVideo
🎯 What it does: Proposes the ConsistentCity two-stage framework: first, it uses the Semantic Flow-guided Diffusion Transformer (SF-DiT) to generate temporally consistent 3D occupancy grids from continuous BEV semantic maps, and then projects them into 2D semantic/depth maps as ControlNet control, ultimately achieving temporally consistent driving scene video synthesis.
ConsNoTrainLoRA: Data-driven Weight Initialization of Low-rank Adapters using Constraints
Debasmit Das (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
GenerationDomain AdaptationImageText
🎯 What it does: The paper proposes a constraint-based, untrained LoRA weight initialization method called CNTLoRA to address the domain shift problem.
Constraint-Aware Feature Learning for Parametric Point Cloud
Xi Cheng (Tsinghua University), Long Zeng (Tsinghua University)
ClassificationRecognitionMultimodalityPoint Cloud
🎯 What it does: This paper proposes a constraint-aware feature learning method for parameterized point clouds, utilizing CAD constraint information to enhance shape recognition and robustness.
Constructing Ophthalmic MLLM for Positioning-diagnosis Collaboration Through Clinical Cognitive Chain Reasoning
Xinyao Liu (Shanghai Artificial Intelligence Laboratory), Diping Song (Shanghai Artificial Intelligence Laboratory)
Object DetectionSegmentationGenerationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageMultimodality
🎯 What it does: Constructed a dedicated multimodal large language model for fundus imaging, FundusExpert, and achieved location-diagnosis collaborative reasoning through FundusGen data automatically generated by Fundus-Engine;
ConstStyle: Robust Domain Generalization with Unified Style Transformation
Nam Duong Tran (Hanoi University of Science and Technology), My T. Thai (University of Florida)
RetrievalDomain AdaptationImage
🎯 What it does: The ConstStyle framework is proposed to reduce domain differences and enhance domain generalization ability by mapping training and testing samples to a unified domain.
Contact-Aware Amodal Completion for Human-Object Interaction via Multi-Regional Inpainting
Seunggeun Chi (Purdue University), Kwonjoon Lee (Honda Research Institute)
Image TranslationRestorationObject DetectionVision Language ModelDiffusion modelImage
🎯 What it does: The paper proposes a framework for completing invisible parts in HOI scenarios based on physical priors and multi-region filling.
Contact-Aware Refinement of Human Pose Pseudo-Ground Truth via Bioimpedance Sensing
Maria-Paola Forte (Max Planck Institute for Intelligent Systems), Michael J. Black (Max Planck Institute for Intelligent Systems)
Pose EstimationOptimizationVideoMultimodality
🎯 What it does: This paper proposes the BioTUCH framework, which combines a wearable wrist bioelectrical impedance sensor with visual 3D joint estimation to automatically detect self-contact and optimize arm posture when contact is detected, thereby improving the reconstruction accuracy of self-contact poses in monocular video.
Context Guided Transformer Entropy Modeling for Video Compression
Junlong Tong (Shanghai Jiao Tong University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative, Institute of Digital Twin, EIT)
CompressionTransformerVideo
🎯 What it does: Proposes the Context Guided Transformer (CGT) conditional entropy model to improve the accuracy and efficiency of entropy modeling in video compression.
Context-Aware Academic Emotion Dataset and Benchmark
Luming Zhao (Zhejiang Gongshang University), Wenwu Yang (Zhejiang Gongshang University)
RecognitionTransformerPrompt EngineeringVision Language ModelVideoBenchmark
🎯 What it does: This study constructed the first academic emotion video dataset RAER in a real-world learning environment and proposed a context-aware emotion recognition framework CLIP-CAER based on CLIP.
ContextFace: Generating Facial Expressions from Emotional Contexts
Min-jung Kim (Korea University), Seung Jun Baek (Korea University)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Proposes ContextFace, a multimodal large language model capable of generating 3D facial expressions based on complex contexts.
Continual Adaptation: Environment-Conditional Parameter Generation for Object Detection in Dynamic Scenarios
Deng Li (Tianjin University), Yahong Han (Tianjin University)
Object DetectionDomain AdaptationComputational EfficiencyDiffusion modelImage
🎯 What it does: This study investigates an adaptive method for object detection implemented through parameter generation in dynamic environments during continuous testing.
Continual Multiple Instance Learning with Enhanced Localization for Histopathological Whole Slide Image Analysis
Byung Hyun Lee (Seoul National University), Se Young Chun (Seoul National University)
ClassificationSegmentationTransformerSupervised Fine-TuningImageMagnetic Resonance Imaging
🎯 What it does: This study proposes a continuous multi-instance learning framework called CoMEL for pathological analysis of whole slide images, balancing bag-level classification and instance-level localization while maintaining a low forgetting rate in continuously incoming tasks.
Continual Personalization for Diffusion Models
Yu-Chien Liao (National Yang Ming Chiao Tung University), Yu-Chiang Frank Wang (National Yang Ming Chiao Tung University)
GenerationData SynthesisOptimizationComputational EfficiencyTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: By evaluating the importance of neurons in the cross-attention layers of diffusion models and filtering them, a concept-related neuron mask is constructed, and only these neurons are fine-tuned, thereby achieving gradual personalization while maintaining the zero-shot capability of the pre-trained model.
Continuous-Time Human Motion Field from Event Cameras
Ziyun Wang (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
Pose EstimationConvolutional Neural NetworkRecurrent Neural NetworkContrastive LearningOptical FlowVideo
🎯 What it does: Utilizing the high temporal resolution of event cameras, we directly predict continuous 3D human motion fields from event streams, generating human poses that can be queried at any moment.
ContraGS: Codebook-Condensed and Trainable Gaussian Splatting for Fast, Memory-Efficient Reconstruction
Sankeerth Durvasula (University of Toronto), Nandita Vijaykumar (University of Toronto)
CompressionOptimizationComputational EfficiencyGaussian SplattingImage
🎯 What it does: This paper proposes ContraGS, a compressible and trainable 3D Gaussian scattering model that significantly reduces memory usage during training and accelerates training and rendering without decreasing the number of Gaussians.
Contrastive Flow Matching
George Stoica (Georgia Tech), Judy Hoffman (Georgia Tech)
GenerationFlow-based ModelContrastive LearningImage
🎯 What it does: This paper proposes a contrastive extension of the flow matching training objective—Contrastive Flow Matching (∆FM), which enhances the discriminability and diversity of conditional generative models by applying contrastive loss to flow vectors.
Contrastive Test-Time Composition of Multiple LoRA Models for Image Generation
Tuna Han Salih Meral (Virginia Tech), Pinar Yanardag (Google)
GenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes a method for seamless integration of multiple LoRA models during the inference phase through contrastive learning on LoRA attention maps, thereby generating high-quality images that simultaneously contain multiple concepts.
Controllable 3D Outdoor Scene Generation via Scene Graphs
Yuheng Liu (Texas A&M University), Ming-Hsuan Yang (University of California Merced)
GenerationData SynthesisAutonomous DrivingGraph Neural NetworkDiffusion modelPoint CloudGraph
🎯 What it does: A 3D outdoor scene generation framework based on scene graph control is proposed, which includes an interactive system, BEV embedded graph, and conditional diffusion model, enabling the automatic generation of user-editable scene graphs into complete 3D urban landscapes.
Controllable and Expressive One-Shot Video Head Swapping
Chaonan Ji (Alibaba Group), Liefeng Bo (Alibaba Group)
Image TranslationGenerationDiffusion modelVideo
🎯 What it does: A one-time video head swapping framework based on diffusion models is proposed, capable of transferring static portrait heads to dynamic videos while keeping the background and body unchanged.
Controllable Feature Whitening for Hyperparameter-Free Bias Mitigation
Yooshin Cho (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationImage
🎯 What it does: A controllable feature whitening framework is proposed to remove the linear correlation between target features and bias features, achieving bias mitigation without hyperparameters.
Controllable Latent Space Augmentation for Digital Pathology
Sofiène Boutaj (CentraleSupélec), Stergios Christodoulidis (CentraleSupélec)
GenerationData SynthesisComputational EfficiencyTransformerImageBiomedical Data
🎯 What it does: A controllable and efficient image enhancement method called HistAug is proposed for digital pathology multi-instance learning.
Controllable Weather Synthesis and Removal with Video Diffusion Models
Chih-Hao Lin (NVIDIA), Zan Gojcic (NVIDIA)
RestorationGenerationData SynthesisVision Language ModelDiffusion modelVideo
🎯 What it does: We propose WEATHERWEAVER, a dual-model framework that can synthesize various weather effects in real videos and accurately remove weather, supporting intensity control for six effects: rain, snow, fog, clouds, puddles, and snow cover.
Controllable-LPMoE: Adapting to Challenging Object Segmentation via Dynamic Local Priors from Mixture-of-Experts
Yanguang Sun (Nanjing University of Science and Technology), Lei Luo (Nankai University)
Object DetectionSegmentationTransformerMixture of ExpertsImage
🎯 What it does: Proposes the Controllable-LPMoE method, which utilizes dynamic local priors and a bidirectional interaction adapter with a frozen large pre-trained model to achieve efficient binary object segmentation.
Controlling Multimodal LLMs via Reward-guided Decoding
Oscar Mañas (Mila - Quebec Artificial Intelligence Institute), Aishwarya Agrawal (Mila - Quebec Artificial Intelligence Institute)
Object DetectionGenerationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a decoding method based on a multimodal reward model (MRGD) to dynamically control the visual accuracy and recall rate of multimodal large language models (MLLMs) during inference, and achieves adjustable computational budgets through search-based decoding.
Cooperative Pseudo Labeling for Unsupervised Federated Classification
Kuangpu Guo (University of Science and Technology of China), Ran He (Institute of Automation, Chinese Academy of Sciences)
ClassificationFederated LearningPrompt EngineeringContrastive LearningImage
🎯 What it does: A framework called FedCoPL is proposed for classification in unlabeled federated learning using CLIP, combining collaborative pseudo-label generation and local prompt aggregation.
CoopTrack: Exploring End-to-End Learning for Efficient Cooperative Sequential Perception
Jiaru Zhong (Tsinghua University), Haibao Yu (University of Hong Kong)
Object TrackingAutonomous DrivingTransformerPoint CloudSequential
🎯 What it does: An end-to-end cooperative 3D multi-object tracking framework called CoopTrack is proposed, achieving a complete pipeline of instance-level collaborative learning and post-decoding fusion.
Coordinate-based Speed of Sound Recovery for Aberration-Corrected Photoacoustic Computed Tomography
Tianao Li (Northwestern University), Emma Alexander (Northwestern University)
RestorationOptimizationComputational EfficiencyBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a self-supervised joint recovery method that can simultaneously recover the initial pressure image and speed field (SOS) from the measurement signals of photoacoustic computed tomography (PACT), achieving end-to-end gradient backpropagation through a differentiable physical model.
CopyrightShield: Enhancing Diffusion Model Security Against Copyright Infringement Attacks
Zhixiang Guo (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
GenerationData SynthesisOptimizationAdversarial AttackDiffusion modelImage
🎯 What it does: A complete defense framework named CopyrightShield is proposed, which first detects potential backdoor samples through spatial masking and data attribution, and then reduces the model's memory of infringing features through adaptive optimization training, thereby preventing copyright infringement attacks.
CoralSRT: Revisiting Coral Reef Semantic Segmentation by Feature Rectification via Self-supervised Guidance
Ziqiang Zheng (Hong Kong University of Science and Technology), Sai-Kit Yeung (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationContrastive LearningImage
🎯 What it does: A coral reef semantic segmentation method based on a self-supervised feature correction module (CoralSRT) is proposed, which can achieve the conversion from sparse points to dense masks through self-supervised feature space correction and label propagation without the need for manual annotation, model fine-tuning, or domain-specific data.
CorrCLIP: Reconstructing Patch Correlations in CLIP for Open-Vocabulary Semantic Segmentation
Dengke Zhang (South China University of Technology), Quan Tang (Pengcheng Laboratory)
SegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes CorrCLIP, which enhances the performance of CLIP in open vocabulary semantic segmentation through reconstructing patch correlations, feature refinement, and correcting segmentation maps.
Correspondence as Video: Test-Time Adaption on SAM2 for Reference Segmentation in the Wild
Haoran Wang (Nanjing University), Yinghuan Shi (Nanjing University)
SegmentationDomain AdaptationMeta LearningDiffusion modelImageVideoBenchmark
🎯 What it does: A CAV-SAM framework is proposed that treats the correspondence between reference images and target images as pseudo-video, utilizing SAM2 for lightweight adaptation during testing.
Correspondence-Free Fast and Robust Spherical Point Pattern Registration
Anik Sarker (Virginia Tech), Alan T. Asbeck (Virginia Tech)
Pose EstimationComputational EfficiencyImagePoint Cloud
🎯 What it does: This paper proposes a correspondence-free spherical point pattern registration algorithm that can quickly estimate the rotational relationship between two sets of spherical point clouds.
Corvid: Improving Multimodal Large Language Models Towards Chain-of-Thought Reasoning
Jingjing Jiang (Shanghai Jiao Tong University), Jun Luo (Nanyang Technological University)
OptimizationConvolutional Neural NetworkTransformerLarge Language ModelContrastive LearningTextMultimodalityChain-of-Thought
🎯 What it does: This paper presents Corvid, a multimodal large language model specifically optimized for chain-of-thought (CoT) reasoning, incorporating a self-verification strategy during the reasoning phase to avoid over- or under-reasoning.
CoSMIC: Continual Self-supervised Learning for Multi-Domain Medical Imaging via Conditional Mutual Information Maximization
Yihang Liu (Tongji University), Heng Tao Shen (Tongji University)
Domain AdaptationRepresentation LearningTransformerContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the CoSMIC framework, which combines continuous self-supervised learning and conditional mutual information maximization to achieve unified pre-training of multi-domain medical images.
COSMO: Combination of Selective Memorization for Low-cost Vision-and-Language Navigation
Siqi Zhang (Tongji University), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
TransformerVision Language ModelMultimodality
🎯 What it does: The COSMO architecture is proposed, combining selective memory state space modules with Transformer to achieve visual language navigation tasks.
CoST: Efficient Collaborative Perception From Unified Spatiotemporal Perspective
Zongheng Tang (Beihang University), Si Liu (Beihang University)
Autonomous DrivingComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: The CoST framework is proposed to achieve unified collaborative perception for multiple vehicles at multiple time points, efficiently fusing information in spatial and temporal dimensions through the STT and USTF modules.
COSTARR: Consolidated Open Set Technique with Attenuation for Robust Recognition
Ryan Rabinowitz (University of Colorado), Terrance E. Boult (University of Colorado)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: A new open set recognition method called COSTARR is proposed, which determines whether a sample belongs to a known category by combining pre-decayed features (original deep features) and post-decayed features (Hadamard product) along with normalization of the values.
CoStoDet-DDPM: Collaborative Training of Stochastic and Deterministic Models Improves Surgical Workflow Anticipation and Recognition
Kaixiang Yang (Wuhan National Laboratory for Optoelectronics), Zhiwei Wang (Wuhan National Laboratory for Optoelectronics)
RecognitionConvolutional Neural NetworkRecurrent Neural NetworkDiffusion modelVideo
🎯 What it does: A collaborative framework (CoStoDet-DDPM) is proposed to simultaneously train a deterministic model and a denoising diffusion probabilistic model (DDPM) for the prediction and recognition of surgical workflows.
CoTMR: Chain-of-Thought Multi-Scale Reasoning for Training-Free Zero-Shot Composed Image Retrieval
Zelong Sun (Renmin University of China), Zhiwu Lu (Renmin University of China)
RetrievalTransformerLarge Language ModelVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: A training-agnostic zero-shot combined image retrieval framework CoTMR is proposed, which integrates Chain-of-Thought Reasoning (CIRCoT) and multi-scale reasoning, directly utilizing large visual language models for unified understanding and reasoning.
CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos
Nikita Karaev (Meta AI), Christian Rupprecht (Visual Geometry Group, University of Oxford)
Object TrackingTransformerVideo
🎯 What it does: A simplified point tracker CoTracker3 has been designed and implemented, and a lightweight unsupervised training scheme using multi-teacher pseudo-labels has been proposed.
CounterPC: Counterfactual Feature Realignment for Unsupervised Domain Adaptation on Point Clouds
Feng Yang (Southeast University), Xuanpeng Li (Southeast University)
ClassificationDomain AdaptationPoint Cloud
🎯 What it does: This paper proposes an unsupervised point cloud domain adaptation framework called CounterPC, which achieves reconstruction and realignment of target domain features through counterfactual interventions on residual features, thereby improving cross-domain classification performance.
Counting Stacked Objects
Corentin Dumery (École Polytechnique Fédérale de Lausanne), Pascal Fua (École Polytechnique Fédérale de Lausanne)
Object DetectionDepth EstimationGaussian SplattingImage
🎯 What it does: This paper proposes a method for counting stacked, low-visibility 3D objects, capable of estimating the total quantity from multi-view images.
CountSE: Soft Exemplar Open-set Object Counting
Shuai Liu (Xi'an Jiaotong University), Wei Ke (Xi'an Jiaotong University)
Object DetectionTransformerVision Language ModelImage
🎯 What it does: A zero-shot object counting method based on soft examples, CountSE, is proposed, which enhances zero-shot counting performance by automatically selecting soft examples in multi-scale features through semantic guidance and filtering with clustering.
Coupling the Generator with Teacher for Effective Data-Free Knowledge Distillation
Xu Chen (Tianjin University), Jialie Shen (City St George's University)
GenerationKnowledge DistillationFlow-based ModelAuto EncoderImage
🎯 What it does: This paper proposes CPNet, which designs the generator as the inverse transformation of the teacher model and couples it with the teacher to form an autoencoder for knowledge distillation without data.
COVTrack: Continuous Open-Vocabulary Tracking via Adaptive Multi-Cue Fusion
Zekun Qian (Tianjin University), Wei Feng (City University of Hong Kong)
Object TrackingKnowledge DistillationVideo
🎯 What it does: This paper proposes the C-TAO dataset with continuous annotations and builds the COVTrack framework based on it for open vocabulary multi-object tracking.
Cracking Instance Jigsaw Puzzles: An Alternative to Multiple Instance Learning for Whole Slide Image Analysis
Xiwen Chen (Clemson University), Aristeidis Sotiras (Washington University in St. Louis)
ClassificationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper proposes replacing traditional multi-instance learning by learning to recover the disrupted instance order (i.e., solving the instance puzzle task) to achieve classification and survival prediction of whole slide images (WSI).
CRAM: Large Scale Video Continual Learning with Bootstrapped Compression
Shivani Mall (Visual Geometry Group, University of Oxford), Joao F. Henriques (Visual Geometry Group, University of Oxford)
ClassificationCompressionVideo
🎯 What it does: This paper proposes a continuous video learning framework called CRAM based on compressed encoding, which can continuously learn action classification from long videos under limited memory.
CreatiLayout: Siamese Multimodal Diffusion Transformer for Creative Layout-to-Image Generation
Hui Zhang (Fudan University), Yu-Gang Jiang (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a SiamLayout framework based on a multimodal diffusion transformer, achieving high-quality controllable generation from layout to image.
Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLMs
Xinyu Fang (Zhejiang University), Dahua Lin (Zhejiang University)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Created Creation-MMBench to evaluate the performance of multimodal large language models in visual creative tasks, including 765 test cases, 51 fine-grained tasks, and providing a text version Creation-MMBench-TO.
Cross-Architecture Distillation Made Simple with Redundancy Suppression
Weijia Zhang (Shanghai Jiao Tong University), Chao Ma (Shanghai Jiao Tong University)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A cross-architecture knowledge distillation method called RSD is proposed, which extracts architecture-independent knowledge through redundancy suppression and transfers it to the student network.
Cross-Category Subjectivity Generalization for Style-Adaptive Sketch Re-ID
Zechao Hu (Wuhan University), Yixiong Zou (Huazhong University of Science and Technology)
RecognitionRetrievalDomain AdaptationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: An Adaptive Incremental Prompt-tuning (AIP) framework is proposed for cross-category subjective style generalization in sketch-based person re-identification (Sketch-based Person Re-ID).
Cross-Granularity Online Optimization with Masked Compensated Information for Learned Image Compression
Haowei Kuang (Peking University), Jiaying Liu (Peking University)
CompressionOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a Cross-Granularity Online Optimization strategy that combines global (image-level) and local (detail-level) online gradient optimization with sparse compensation to enhance the rate-distortion performance of learned image compression.
Cross-modal Ship Re-Identification via Optical and SAR Imagery: A Novel Dataset and Method
Han Wang (Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences), Zhuang Zhou (Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences)
RecognitionRetrievalTransformerContrastive LearningImageMultimodality
🎯 What it does: A cross-modal ship re-identification dataset HOSS ReID has been established, and a cross-modal re-identification method based on Vision Transformer, TransOSS, has been proposed.
Cross-Subject Mind Decoding from Inaccurate Representations
Yangyang Xu (Harbin Institute of Technology), Tingting Zhu (Singapore Management University)
RestorationGenerationTransformerDiffusion modelAuto EncoderImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: A cross-subject brain image decoding framework is proposed, capable of recovering original visual stimulus images from fMRI signals.
Cross-View Isolated Sign Language Recognition via View Synthesis and Feature Disentanglement
Xin Shen (University of Queensland), Xin Yu (University of Queensland)
RecognitionData SynthesisContrastive LearningVideo
🎯 What it does: A new two-stage framework is proposed for cross-view isolated sign language recognition (CV-ISLR), which includes view synthesis and contrastive multi-task view-semantic recognition.
CryoFastAR: Fast Cryo-EM Ab initio Reconstruction Made Easy
Jiakai Zhang (ShanghaiTech University), Jingyi Yu (ShanghaiTech University)
Pose EstimationComputational EfficiencyProtein Structure PredictionTransformerSupervised Fine-TuningImage
🎯 What it does: This work proposes CryoFastAR, a fast cryo-electron microscopy (cryo-EM) ab initio reconstruction method based on a geometric foundational model, capable of predicting poses and directly performing three-dimensional reconstruction from a large number of disordered, unaligned, and highly noisy particle images in one go, avoiding traditional iterative optimization processes.
CSD-VAR: Content-Style Decomposition in Visual Autoregressive Models
Quang-Binh Nguyen (Qualcomm AI Research), Khoi Nguyen (Qualcomm AI Research)
Image TranslationGenerationTransformerImage
🎯 What it does: Using a Visual Autoregressive Model (VAR) to achieve content-style decomposition (CSD) of a single image, enabling independent control of content reconstruction and style transfer.
CT-ScanGaze: A Dataset and Baselines for 3D Volumetric Scanpath Modeling
Trong Thang Pham (University of Arkansas), Ngan Le (University of Liverpool)
Explainability and InterpretabilityTransformerBiomedical DataComputed Tomography
🎯 What it does: This paper presents the first publicly available CT eye movement dataset, CT-ScanGaze, and proposes a 3D eye movement path prediction model, CT-Searcher, which further opens new directions for explainable AI in medical imaging.
CULTURE3D: A Large-Scale and Diverse Dataset of Cultural Landmarks and Terrains for Gaussian-Based Scene Rendering
Xinyi Zheng (University of Bristol), Junxiao Shen (University of Bristol)
GenerationData SynthesisOptimizationGaussian SplattingImagePoint CloudMeshBenchmark
🎯 What it does: This paper constructs a large-scale, ultra-high-resolution cultural heritage scene dataset called CULTURE3D and benchmarks various Gaussian point-based 3D reconstruction methods.
CuMPerLay: Learning Cubical Multiparameter Persistence Vectorizations
Caner Korkmaz (Imperial College London), Tolga Birdal (Imperial College London)
ClassificationSegmentationTransformerImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes CuMPerLay, a differentiable multi-parameter persistence vectorization layer that embeds cubic multi-parameter persistence (CMP) into deep learning networks, enabling end-to-end learning of topological feature extraction.
CuRe: Cultural Gaps in the Long Tail of Text-to-Image Systems
Aniket Rege (University of Wisconsin Madison), Ramya Korlakai Vinayak (University of Wisconsin Madison)
GenerationData SynthesisVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: A CuRe benchmark and scoring system is proposed to evaluate the representativeness of text-to-image models in cross-cultural performance.
Curve-Aware Gaussian Splatting for 3D Parametric Curve Reconstruction
Zhirui Gao (National University of Defense Technology), Kai Xu (National University of Defense Technology)
OptimizationGaussian SplattingPoint Cloud
🎯 What it does: An end-to-end single-stage framework is proposed, which directly optimizes three-dimensional parameter curves from multi-view edge images, avoiding the error accumulation of traditional two-stage methods.
Customizing Domain Adapters for Domain Generalization
Yuyang Ji (University of Wisconsin Madison), Yong Jae Lee (University of Illinois Urbana Champaign)
Domain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Custom Domain Adapter (CDA) framework for domain generalization, utilizing lightweight ViT and CNN adapters to focus on learning the data features of their respective domains, and dynamically merging predictions through a domain router;
CutS3D: Cutting Semantics in 3D for 2D Unsupervised Instance Segmentation
Leon Sick (Ulm University), Timo Ropinski (Ulm University)
Object DetectionSegmentationTransformerImagePoint Cloud
🎯 What it does: A method for 2D unsupervised instance segmentation using 3D information is proposed, which first uses depth point clouds to cut semantic masks in 3D space to obtain pseudo-instance masks, and then trains a class-agnostic detector.
CVFusion: Cross-View Fusion of 4D Radar and Camera for 3D Object Detection
Hanzhi Zhong (Zhejiang University), Eryun Liu (Zhejiang University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkTransformerImageMultimodalityPoint Cloud
🎯 What it does: This paper designs a two-stage cross-view fusion network called CVFusion, which fuses 4D radar point clouds with camera images for 3D object detection.
CVPT: Cross Visual Prompt Tuning
Lingyun Huang (Hunan University), Yaonan Wang (Hunan University)
ClassificationSegmentationTransformerPrompt EngineeringImage
🎯 What it does: A Cross Visual Prompt Tuning (CVPT) method is proposed, which improves prompt tuning in visual tasks through a cross-attention module, decoupling prompts from the self-attention of input tokens, achieving more efficient and powerful parameter-efficient fine-tuning.
CWNet: Causal Wavelet Network for Low-Light Image Enhancement
Tongshun Zhang (Jilin University), Qiuzhan Zhou (Jilin University)
RestorationContrastive LearningImage
🎯 What it does: A low-light image enhancement network called CWNet based on causal reasoning and wavelet transform is proposed.
Cycle Consistency as Reward: Learning Image-Text Alignment without Human Preferences
Hyojin Bahng (Massachusetts Institute of Technology), Phillip Isola (Massachusetts Institute of Technology)
GenerationOptimizationTransformerReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Constructing unsupervised preference labels through cycle consistency (image→text→image or text→image→text) to train the reward model CycleReward for evaluating and optimizing image-text alignment.
Cycle-Consistent Learning for Joint Layout-to-Image Generation and Object Detection
Xinhao Cai (Nanjing University of Science and Technology), Wenguan Wang (Zhejiang University)
Image TranslationObject DetectionGenerationDiffusion modelImage
🎯 What it does: The GDCC (Generation-Detection Cycle-Consistent) framework is proposed, which simultaneously optimizes the layout-to-image (L2I) and object detection (OD) tasks in an end-to-end training process.
CycleVAR: Repurposing Autoregressive Model for Unsupervised One-Step Image Translation
Yi Liu (Beihang University), Si Liu
Image TranslationGenerationGenerative Adversarial NetworkImage
🎯 What it does: The CycleVAR framework is proposed, which applies Visual Autoregressive Models (VAR) for unsupervised single-step image translation and achieves end-to-end training.
D-Attn: Decomposed Attention for Large Vision-and-Language Model
Chia-Wen Kuo (ByteDance Intelligent Creation), Longyin Wen (ByteDance Intelligent Creation)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: The Decomposed Attention (D-Attn) architecture is proposed, which splits the self-attention in LVLM into three parts: visual-visual, text-visual, and text-text, and fuses them through α-weighting to achieve flexible processing of visual tokens.
D2ST-Adapter: Disentangled-and-Deformable Spatio-Temporal Adapter for Few-shot Action Recognition
Wenjie Pei (Harbin Institute of Technology), Jun Yu (Harbin Institute of Technology)
RecognitionTransformerSupervised Fine-TuningVideo
🎯 What it does: Designed and implemented a dual-path video adapter based on an adapter (D ST-Adapter 2) for efficiently transferring from pre-trained image models to the video domain in few-shot action recognition tasks.
D3: Training-Free AI-Generated Video Detection Using Second-Order Features
Chende Zheng (Xi'an Jiaotong University), Chao Shen (City University of Hong Kong)
Object DetectionAnomaly DetectionOptical FlowVideo
🎯 What it does: This study investigates the differences in second-order temporal features (acceleration) between AI-generated videos and real videos, proposing a training-free detection method called D3 based on second-order central differences.
D3QE: Learning Discrete Distribution Discrepancy-aware Quantization Error for Autoregressive-Generated Image Detection
Yanran Zhang (Tsinghua University), Jiwen Lu (Tsinghua University)
GenerationAnomaly DetectionTransformerDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A method based on discrete distribution difference quantization error (D QE 3) is proposed to detect images generated by visual autoregressive models.
DAA*: Deep Angular A Star for Image-based Path Planning
Zhiwei Xu (University of Melbourne)
Autonomous DrivingOptimizationConvolutional Neural NetworkImageVideo
🎯 What it does: A deep angle A* (DAA*) method is proposed to improve image-based path planning by learning the Path Angle Freedom (PAF), making the predicted paths closer to expert demonstrations and achieving better smoothness.
DACoN: DINO for Anime Paint Bucket Colorization with Any Number of Reference Images
Kazuma Nagata (Tokyo Denki University), Naoshi Kaneko (Tokyo Denki University)
Image TranslationGenerationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: DACoN is designed to colorize animation line art using DINOv2 and U-Net with multiple reference images, taking into account both key frames and continuous frames.
DADet: Safeguarding Image Conditional Diffusion Models against Adversarial and Backdoor Attacks via Diffusion Anomaly Detection
Hongwei Yu (University of Science and Technology Beijing), Jiansheng Chen (University of Science and Technology Beijing)
Anomaly DetectionAdversarial AttackDiffusion modelImage
🎯 What it does: This paper studies the vulnerability of image conditional diffusion models to backdoor attacks and adversarial attacks, and proposes a defense method based on Diffusion Anomaly Detection (DADet).
DADM: Dual Alignment of Domain and Modality for Face Anti-spoofing
Jingyi Yang (University of Science and Technology of China), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)
RecognitionDomain AdaptationTransformerContrastive LearningImageMultimodality
🎯 What it does: A dual-alignment multimodal domain adaptation (DADM) method is proposed to enhance the cross-domain and cross-modal generalization ability of facial attack detection (FAS).
DALIP: Distribution Alignment-based Language-Image Pre-Training for Domain-Specific Data
Junjie Wu (Tianjin University), Sen Xu (Dalian University of Technology)
ClassificationDomain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkAgriculture Related
🎯 What it does: This paper proposes a distribution alignment-based CLIP pre-training method (DALIP), which aligns the model using the first and second-order statistics of image-text feature distributions, and constructs the PlantMix-13M plant-specific dataset to further enhance the model's performance in the fine-grained biological domain while maintaining compatibility with the general domain.
DAMap: Distance-aware MapNet for High Quality HD Map Construction
Jinpeng Dong (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
Autonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes an online HD map construction framework based on Transformer, named DAMap, which focuses on addressing performance bottlenecks caused by task label mismatches and task feature sharing in high-quality predictions.
DanceEditor: Towards Iterative Editable Music-driven Dance Generation with Open-Vocabulary Descriptions
Hengyuan Zhang (Peking University), Sirui Han (Hong Kong University of Science and Technology)
GenerationTransformerVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: A music-driven dance generation framework called DanceEditor is proposed, which allows for iterative and editable dance generation. It supports users in adjusting dance movements using open vocabulary text descriptions through multiple rounds of editing.
DAP-MAE: Domain-Adaptive Point Cloud Masked Autoencoder for Effective Cross-Domain Learning
Ziqi Gao (Shenzhen University), Linlin Shen (Shenzhen University)
ClassificationObject DetectionSegmentationDomain AdaptationTransformerAuto EncoderContrastive LearningPoint Cloud
🎯 What it does: A Domain Adaptive Point Cloud Masked Autoencoder (DAP-MAE) is proposed, which can fully utilize multi-domain point cloud data during a single pre-training phase and achieve high performance on various downstream tasks (classification, segmentation, expression recognition, detection).
Dark-ISP: Enhancing RAW Image Processing for Low-Light Object Detection
Jiasheng Guo (Fudan University), Jian Pu (Fudan University)
Object DetectionImage
🎯 What it does: A lightweight adaptive RAW processing plugin Dark-ISP is proposed for object detection in low-light environments.
DASH: 4D Hash Encoding with Self-Supervised Decomposition for Real-Time Dynamic Scene Rendering
Jie Chen (University of Science and Technology of China), Xiaoyan Sun (Institute of Artificial Intelligence, Hefei Comprehensive National Science Center)
GenerationComputational EfficiencyNeural Radiance FieldVideo
🎯 What it does: A real-time dynamic scene rendering framework named DASH is proposed, utilizing 4D hash encoding and self-supervised dynamic-static decomposition to achieve high-quality, real-time rendering;
DASH: Detection and Assessment of Systematic Hallucinations of VLMs
Maximilian Augustin (Tuebingen AI Center University of Tuebingen), Matthias Hein (Tuebingen AI Center University of Tuebingen)
Object DetectionRetrievalOptimizationTransformerLarge Language ModelVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: This paper proposes an automated large-scale pipeline named DASH (Detection and Assessment of Systematic Hallucinations) for identifying systematic 'false positive' hallucinations of visual language models (VLM) on open-world images and clustering them into semantically similar image clusters.
DATA: Domain-And-Time Alignment for High-Quality Feature Fusion in Collaborative Perception
Chengchang Tian (Southeast University), Wei Hong (Southeast University)
Domain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: This paper studies the domain gap caused by hardware diversity and deployment differences in collaborative perception, as well as the temporal mismatch issues caused by communication delays. It proposes a framework named DATA, aimed at achieving domain alignment, temporal alignment, and feature aggregation during the feature acquisition phase to enhance multi-agent perception performance.
Dataset Distillation as Data Compression: A Rate-Utility Perspective
Youneng Bao (City University of Hong Kong), Kede Ma (City University of Hong Kong)
CompressionKnowledge DistillationImage
🎯 What it does: A framework is proposed in dataset distillation that unifies the optimization of compression rate and distillation effectiveness.
Dataset Distillation via the Wasserstein Metric
Haoyang Liu (University of Illinois at Urbana-Champaign), Haohan Wang (University of Illinois at Urbana-Champaign)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a dataset distillation method based on Wasserstein barycenter (WMDD), which optimizes synthetic samples in the feature space of a pre-trained model to align their distribution with the Wasserstein barycenter of the real data.
Dataset Distillation via Vision-Language Category Prototype
Yawen Zou (University of Toyama), Chao Zhang (University of Toyama)
ClassificationData SynthesisKnowledge DistillationTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
🎯 What it does: By introducing visual-language prototypes (image prototypes + text prototypes) into dataset distillation, a more semantically consistent and higher-performing synthetic dataset is generated.
Dataset Ownership Verification for Pre-trained Masked Models
Yuechen Xie (Zhejiang University), Mingli Song (Zhejiang University)
TransformerContrastive LearningImageText
🎯 What it does: A black-box dataset ownership verification method DOV4MM based on the difficulty of reconstructing relative embeddings is proposed, specifically targeting pre-trained mask models.
DAViD: Data-efficient and Accurate Vision Models from Synthetic Data
Fatemeh Saleh (Microsoft), Tadas Baltrusaitis (Microsoft)
SegmentationDepth EstimationTransformerImage
🎯 What it does: A human-centered vision model based on high-fidelity synthetic data SynthHuman is proposed, using a single Transformer architecture to simultaneously accomplish relative depth estimation, surface normal estimation, and soft foreground segmentation.