ICCV 2025 Papers — Page 17
IEEE/CVF International Conference on Computer Vision · 2701 papers
MuGS: Multi-Baseline Generalizable Gaussian Splatting Reconstruction
Yaopeng Lou (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
GenerationData SynthesisDepth EstimationGaussian SplattingImage
🎯 What it does: A multi-baseline generalizable 3D Gaussian splatting method, MuGS, is proposed for novel view synthesis, capable of generating high-quality images uniformly under sparse input and different baselines (small and large baselines).
Multi-Cache Enhanced Prototype Learning for Test-Time Generalization of Vision-Language Models
Xinyu Chen (Shanghai University), Ruirui Li (Beijing University of Chemical Technology)
Domain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A multi-cache prototype learning framework (MCP) and its residual enhanced version (MCP++) are proposed, which dynamically construct visual-text prototypes during the inference phase through low-entropy cache, alignment cache, and negative sample cache, thereby achieving adaptive generalization of zero-shot visual-language models.
Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs
Jeongseok Hyun (NAVER Cloud), Minho Shim (NAVER Cloud)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: A training-independent multi-granularity spatiotemporal token merging method (STTM) is proposed in video large language models to reduce the number of visual tokens and improve inference speed.
Multi-identity Human Image Animation with Structural Video Diffusion
Zhenzhi Wang (Chinese University of Hong Kong), Bo Dai (University of Hong Kong)
GenerationData SynthesisPose EstimationDepth EstimationDiffusion modelVideo
🎯 What it does: This paper proposes a structured video diffusion framework for generating multi-identity interactive videos from a single portrait.
Multi-Modal Few-Shot Temporal Action Segmentation
Zijia Lu (Northeastern University), Ehsan Elhamifar (Northeastern University)
RecognitionSegmentationGraph Neural NetworkTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes the Multi-modal Few-shot Temporal Action Segmentation (MMF-TAS) task and achieves adaptive segmentation for the new task through a Prototype Graph Network.
Multi-modal Identity Extraction
Ryan Webster (Inria), Teddy Furon (Inria)
Adversarial AttackTransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes and implements an identity extraction attack on multimodal models, capable of recovering personal names from the model using only facial images, extending previous research that only detected identities of training set members.
Multi-modal Multi-platform Person Re-Identification: Benchmark and Method
Ruiyang Ha (ShanghaiTech University), Jingya Wang (Queen Mary University London)
RecognitionRetrievalPrompt EngineeringVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: A multi-modal multi-platform pedestrian re-identification (MP-ReID) benchmark dataset is proposed, along with a unified Prompt learning framework called Uni-Prompt ReID.
Multi-Modal Multi-Task Unified Embedding Model (M3T-UEM): A Task-Adaptive Representation Learning Framework
Rohan Sharma (Amazon), Qingjun Cui (Amazon)
RetrievalRepresentation LearningTransformerLarge Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes the M3T-UEM framework, utilizing LLM backbone to achieve unified embedding for multi-modal and multi-task scenarios, significantly enhancing retrieval and zero-shot classification performance.
Multi-modal Segment Anything Model for Camouflaged Scene Segmentation
Guangyu Ren (Xi'an Jiaotong-Liverpool University), Tania Stathaki (Imperial College London)
Object DetectionSegmentationTransformerVision Language ModelImageMultimodalityBenchmark
🎯 What it does: For target segmentation in concealed scene images, a multi-modal prompt Segment Anything model (MM-SAM) is proposed, which completes segmentation without the need for manual boxes or point prompts.
Multi-Object Sketch Animation by Scene Decomposition and Motion Planning
Jingyu Liu (Renmin University of China), Xirong Li (Renmin University of China)
GenerationOptimizationTransformerLarge Language ModelScore-based ModelImageVideo
🎯 What it does: Proposes MoSketch, an iterative optimization framework for multi-object sketch animation that does not require training data;
Multi-scenario Overlapping Text Segmentation with Depth Awareness
Yang Liu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: Proposes a multi-scene overlapping text segmentation task, constructs a new dataset called MOT, designs a hierarchical synthesis strategy (HSOT), and develops a depth-guided decoder to achieve segmentation.
Multi-Schema Proximity Network for Composed Image Retrieval
Jiangming Shi (Xiamen University), Yanyun Qu (Xiamen University)
RetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A multi-modal proximity network (MAPNet) is proposed for combined image retrieval, generating queries by combining reference images and text descriptions to match target images.
Multi-turn Consistent Image Editing
Zijun Zhou (Institute of Computing Technology, Chinese Academy of Sciences), Fan Tang (Institute of Computing Technology, Chinese Academy of Sciences)
Image TranslationRestorationGenerationTransformerRectified FlowImageOrdinary Differential Equation
🎯 What it does: A multi-round image editing framework is proposed, which achieves precise restoration and stable sampling of images through flow matching inversion and dual-objective LQR control in each editing iteration, and utilizes adaptive attention to guide local edits, ensuring coherence and structural preservation in the edits.
Multi-View 3D Point Tracking
Frano Rajič (ETH Zurich), Siyu Tang (ETH Zurich)
Object TrackingDepth EstimationConvolutional Neural NetworkTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper presents MVTracker, an end-to-end 3D point tracker based on multi-view perspectives, capable of online tracking of arbitrary 3D points in practical scenarios such as with four cameras.
Multi-view Gaze Target Estimation
Qiaomu Miao (Stony Brook University), Dimitris Samaras (Stony Brook University)
Object DetectionPose EstimationDepth EstimationTransformerImage
🎯 What it does: This paper proposes a Multi-View Gaze Target Estimation (GTE) framework based on multiple camera perspectives, which jointly infers the gaze target location using images from two cameras and can still make predictions when the face or target is not visible from a single perspective.
Multi-View Slot Attention Using Paraphrased Texts for Face Anti-Spoofing
Jeongmin Yu (Yonsei University), Ha Young Kim (Yonsei University)
RecognitionDomain AdaptationRecurrent Neural NetworkPrompt EngineeringVision Language ModelImageVideo
🎯 What it does: The MVP-FAS framework is proposed, achieving cross-domain generalization for face spoofing detection through multi-view slot attention and multi-text patch alignment.
MultiADS: Defect-aware Supervision for Multi-type Anomaly Detection and Segmentation in Zero-Shot Learning
Ylli Sadikaj (University of Vienna), Claudia Plant (University of Vienna)
SegmentationAnomaly DetectionTransformerVision Language ModelImage
🎯 What it does: This paper proposes MultiADS, a zero-shot/few-shot multi-type defect detection and segmentation framework that can distinguish multiple defect types at the pixel level.
Multidimensional Byte Pair Encoding: Shortened Sequences for Improved Visual Data Generation
Tim Elsner (RWTH Aachen University), Leif Kobbelt (RWTH Aachen University)
GenerationCompressionTransformerAuto EncoderImage
🎯 What it does: This paper proposes Multi-Dimensional Byte Pair Encoding (MDBPE), which compresses the sequence length of visual data by statistically merging frequent token pairs in two-dimensional (or three-dimensional) space, allowing Transformers to train and infer faster and more efficiently.
MultiModal Action Conditioned Video Simulation
Yichen Li (Massachusetts Institute of Technology), Antonio Torralba (Massachusetts Institute of Technology)
GenerationData SynthesisMixture of ExpertsDiffusion modelVideoMultimodality
🎯 What it does: A fine-grained video generation and simulation framework based on multimodal (haptic, electromyography, hand posture, body posture, gaze) action signals is proposed, and future video frame prediction is achieved through a diffusion model.
Multimodal Large Language Model-Guided ISP Hyperparameter Optimization with Dynamic Preference Learning
Xinyu Sun (Beijing Jiaotong University), Juan Wang (Chinese Academy of Sciences)
Object DetectionSegmentationOptimizationHyperparameter SearchLarge Language ModelReinforcement LearningPrompt EngineeringImageTextMultimodality
🎯 What it does: This paper proposes a hyperparameter optimization framework for ISP based on a multimodal large language model, guiding hyperparameter search through generated text prompts.
Multimodal Latent Diffusion Model for Complex Sewing Pattern Generation
Shengqi Liu (Shanghai Jiao Tong University), Yichao Yan
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: A multi-modal latent diffusion model called SewingLDM is proposed to generate complex sewing patterns conditioned on text, clothing sketches, and body shapes.
Multimodal LLM Guided Exploration and Active Mapping using Fisher Information
Wen Jiang (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
OptimizationRobotic IntelligenceTransformerLarge Language ModelGaussian SplattingSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: This paper proposes an active mapping system based on 3D Gaussian Splatting, utilizing a multimodal LLM for long-term goal planning, and subsequently employs Fisher information-driven path planning to maximize information gain in the short term and control localization error.
Multimodal LLMs as Customized Reward Models for Text-to-Image Generation
Shijie Zhou (University at Buffalo), Changyou Chen (University at Buffalo)
GenerationRecommendation SystemTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A multimodal reward model named LLaVA-Reward was designed and trained to evaluate the performance of text-to-image generation across multiple dimensions, including alignment, authenticity, aesthetics, and safety.
Multimodal Prompt Alignment for Facial Expression Recognition
Fuyan Ma (Chinese Academy of Military Science), Shutao Li (Hunan University)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringImageMultimodality
🎯 What it does: A multi-modal prompt alignment framework MPA-FER is proposed, keeping the CLIP visual encoder frozen, and achieving fine-grained cross-modal learning of facial expressions through soft and hard prompts, visual prompts, and prototype-guided alignment.
Multispectral Demosaicing via Dual Cameras
SaiKiran Tedla (Samsung Electronics), Michael S. Brown (York University)
RestorationConvolutional Neural NetworkOptical FlowImage
🎯 What it does: A deep learning framework for joint demosaicing using dual-camera RGB and multispectral images is proposed.
MultiVerse: A Multi-Turn Conversation Benchmark for Evaluating Large Vision and Language Models
Young-Jun Lee (KAIST), Ho-Jin Choi (KAIST)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes MULTIVERSE—a multi-turn dialogue evaluation benchmark covering 484 sub-tasks, 8 types of interaction goals, and 25 image domains, providing an instantiated checklist evaluation for each round.
MultiverSeg: Scalable Interactive Segmentation of Biomedical Imaging Datasets with In-Context Guidance
Hallee E. Wong (Massachusetts Institute of Technology), Adrian V. Dalca (Massachusetts Institute of Technology)
SegmentationConvolutional Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: MultiverSeg is proposed, a scalable interactive segmentation framework that utilizes user clicks/lines/box prompts and dynamically growing contextual samples to quickly complete full image segmentation of new datasets without retraining the model.
MUNBa: Machine Unlearning via Nash Bargaining
Jing Wu (Monash University), Mehrtash Harandi (Monash University)
GenerationData SynthesisConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelImage
🎯 What it does: MUNBa proposes a closed-form gradient negotiation method by modeling the machine unlearning problem as a two-player Nash game, achieving a balance between forgetting and retention, and addressing gradient conflicts and dominance issues.
MUSE-VL: Modeling Unified VLM through Semantic Discrete Encoding
Rongchang Xie (ByteDance), Chang Liu (ByteDance)
GenerationRetrievalTransformerVision Language ModelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: MUSE-VL is proposed, a unified visual-language model capable of simultaneous visual understanding and image generation.
MUSE: Multi-Subject Unified Synthesis via Explicit Layout Semantic Expansion
Fei Peng (Beijing University of Posts and Telecommunications), Huiyuan Fu (Beijing University of Posts and Telecommunications)
GenerationData SynthesisDiffusion modelImageTextBenchmark
🎯 What it does: The MUSE framework is proposed, which enables multi-agent layout control and identity preservation synthesis in a single inference.
Music Grounding by Short Video
Zijie Xin (Renmin University of China), Xirong Li (Renmin University of China)
RetrievalTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes the music grounding task MGSV, which aims to locate suitable background music segments for short videos in a music library.
Music-Aligned Holistic 3D Dance Generation via Hierarchical Motion Modeling
Xiaojie Li (Tsinghua University), Zhi Wang (ByteDance)
GenerationRetrievalTransformerVideoMultimodality
🎯 What it does: This paper presents the SoulDance full-body dance dataset and the SoulNet generation framework for music-driven 3D full-body dance generation.
MV-Adapter: Multi-View Consistent Image Generation Made Easy
Zehuan Huang (Beihang University), Lu Sheng (Beihang University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes MV-Adapter, a pluggable adapter that enables existing text-to-image diffusion models and their derivatives to generate high-resolution images with consistent multi-view perspectives without altering the original network structure, while supporting conditional guidance from text, images, camera parameters, and geometric information.
MVGBench: a Comprehensive Benchmark for Multi-view Generation Models
Xianghui Xie (University of Tübingen), Gerard Pons-Moll (Max Planck Institute for Informatics)
GenerationData SynthesisVision Language ModelGaussian SplattingImageBenchmark
🎯 What it does: MVGBench is proposed, a unified evaluation framework for multi-view image generation models, covering ten metrics such as 3D consistency, image quality, and semantic consistency.
MVQA: Mamba with Unified Sampling for Efficient Video Quality Assessment
Yachun Mi (Harbin Institute of Technology), Shaohui Liu (Harbin Institute of Technology)
OptimizationComputational EfficiencyConvolutional Neural NetworkVideo
🎯 What it does: An efficient video quality assessment framework MVQA based on the Mamba state space model is designed and implemented, and a Unified Semantic and Distortion Sampling (USDS) method is proposed, which integrates low-resolution semantic information with high-resolution distortion details.
MVTrajecter: Multi-View Pedestrian Tracking with Trajectory Motion Cost and Trajectory Appearance Cost
Taiga Yamane (NTT Corporation), Shota Orihashi (NTT Corporation)
Object TrackingVideoBenchmark
🎯 What it does: An end-to-end multi-view pedestrian tracking method called MVTrajecter is proposed, which utilizes motion and appearance information from multi-frame trajectories to achieve more robust associations.
NAPPure: Adversarial Purification for Robust Image Classification under Non-Additive Perturbations
Junjie Nan (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
ClassificationAutonomous DrivingAdversarial AttackDiffusion modelImage
🎯 What it does: The NAPPure framework is proposed, which purifies non-additive adversarial perturbations by jointly optimizing clean images and perturbation parameters.
NATRA: Noise-Agnostic Framework for Trajectory Prediction with Noisy Observations
Rongqing Li (Beijing Institute of Technology), Jun Zhou (Ant Group)
Autonomous DrivingOptimizationGraph Neural NetworkContrastive LearningTime SeriesSequential
🎯 What it does: This paper proposes the NATRA framework, which can achieve collaborative optimization of trajectory denoising and prediction under arbitrary noise observations.
Nautilus: Locality-aware Autoencoder for Scalable Mesh Generation
Yuxuan Wang (Nanyang Technological University), Hanwang Zhang (Nanyang Technological University)
GenerationTransformerAuto EncoderPoint CloudMesh
🎯 What it does: A scalable mesh generation framework called Nautilus based on autoencoders is proposed, capable of generating high-quality artistic style meshes directly from point clouds or single-view images;
NAVER: A Neuro-Symbolic Compositional Automaton for Visual Grounding with Explicit Logic Reasoning
Zhixi Cai (Monash University), Hamid Rezatofighi (Monash University)
Object DetectionSegmentationLarge Language ModelVision Language ModelImageText
🎯 What it does: This paper presents NAVER, a visual grounding method implemented through a neural-symbolic deterministic finite state automaton; in this method, entities, attributes, and relationships are first extracted using a vision-language model, then the query is converted into a ProbLog logical expression for probabilistic reasoning, and finally, the answer is verified through the vision-language model and self-correction is implemented in the automaton.
NavMorph: A Self-Evolving World Model for Vision-and-Language Navigation in Continuous Environments
Xuan Yao (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
Recurrent Neural NetworkWorld ModelMultimodality
🎯 What it does: This paper proposes NavMorph, a self-evolving world model framework for visual-language navigation (VLN-CE) tasks in continuous environments, capable of dynamically updating environmental representations and performing proactive action planning during online testing.
NavQ: Learning a Q-Model for Foresighted Vision-and-Language Navigation
Peiran Xu (Peking University), Yadong Mu (Peking University)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A self-supervised Q model and future encoder-based prospective visual language navigation agent, NavQ, has been developed to achieve forward decision-making in goal-oriented tasks.
NegRefine: Refining Negative Label-Based Zero-Shot OOD Detection
Amirhossein Ansari (Simon Fraser University), Pulei Xiong (National Research Council Canada)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: This paper proposes an improved zero-shot OOD detection framework called NegRefine, which utilizes negative label filtering and multi-label matching scoring to achieve a more robust distinction between in-distribution and OOD.
Neighboring Autoregressive Modeling for Efficient Visual Generation
Yefei He (Zhejiang University), Bohan Zhuang (Zhejiang University)
GenerationData SynthesisComputational EfficiencyTransformerImageVideoTextBenchmark
🎯 What it does: Proposes Neighborhood Autoregressive Modeling (NAR), treating visual generation as a step-by-step outpainting process starting from an initial token, using a 'next neighbor prediction' mechanism that progresses from near to far.
NeRF Is a Valuable Assistant for 3D Gaussian Splatting
Shuangkang Fang (Beihang University), Shuchang Zhou (StepFun)
GenerationOptimizationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Combining NeRF with 3D Gaussian Splatting, the NeRF-GS framework is proposed, achieving joint optimization of both.
NETracer: A Topology-Aware Iterative Tracing Approach for Tubular Structure Extraction
Chao Liu (Zhejiang University), Nenggan Zheng (Zhejiang University)
Object TrackingSegmentationConvolutional Neural NetworkImage
🎯 What it does: An iterative tracking framework named NETracer is proposed, which utilizes a node-edge estimation network and a geographic distance search strategy to accurately extract tubular structures in images.
NeuFrameQ: Neural Frame Fields for Scalable and Generalizable Anisotropic Quadrangulation
Ying-Tian Liu (Tsinghua University), Song-Hai Zhang (Tsinghua University)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: Proposes the NeuFrameQ framework, which predicts framework fields from point clouds and generates high-quality quadrilateral meshes, achieving end-to-end automation from triangular meshes to quadrilateral meshes.
Neural Architecture Search Driven by Locally Guided Diffusion for Personalized Federated Learning
Peng Liao (East China University of Science and Technology), Han Hu (East China University of Science and Technology)
Federated LearningNeural Architecture SearchDiffusion modelImageStochastic Differential Equation
🎯 What it does: A personalized federated learning framework based on client-side NAS, PerFedSDE-NAS, is proposed, which utilizes a conditional diffusion model to generate candidate architectures and filters them through an RBF proxy, combined with archived training and improved supernetwork aggregation to achieve efficient personalized model search.
Neural Compression for 3D Geometry Sets
Siyu Ren (City University of Hong Kong), Wenping Wang (Texas A&M University)
CompressionPoint CloudMesh
🎯 What it does: This paper proposes the NeCGS framework for compressing collections of multi-class 3D mesh models, which can maintain detail at extremely high compression ratios.
Neural Inverse Rendering for High-Accuracy 3D Measurement of Moving Objects with Fewer Phase-Shifting Patterns
Yuki Urakawa (Institute of Science Tokyo), Yoshihiro Watanabe (Institute of Science Tokyo)
Depth EstimationOptimizationNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper proposes a phase-shift structured light measurement method that integrates displacement fields into a multi-camera single-projector neural inverse rendering framework for high-precision 3D measurement of moving objects.
Neural Multi-View Self-Calibrated Photometric Stereo without Photometric Stereo Cues
Xu Cao (CyberAgent), Takafumi Taketomi (CyberAgent)
RestorationData SynthesisNeural Radiance FieldImage
🎯 What it does: An end-to-end neural inverse rendering framework is proposed, capable of simultaneously recovering geometry, spatially varying BRDF, and light source direction and intensity from multi-view single-light images.
Neural Shell Texture Splatting: More Details and Fewer Primitives
Xin Zhang (Zhejiang University), Weiwei Xu (Zhejiang University)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingMesh
🎯 What it does: This paper proposes NeST-Splatting, which achieves a complete decoupling of geometry and appearance by using 2D Gaussian primitives to describe geometry and mapping texture information to a global neural shell texture (multi-resolution hash grid), maintaining high-quality textures while significantly reducing the number of required primitives.
Neural Solver of Dichromatic Reflection Model for Specular Highlight Removal
Gang Fu (Shenzhen University)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: A neural network-based dichromatic reflection model (DRM) solver is proposed for the removal of specular highlights in a single image.
NeuraLeaf: Neural Parametric Leaf Models with Shape and Deformation Disentanglement
Yang Yang (Osaka University), Fumio Okura (Microsoft Research Asia)
GenerationData SynthesisGenerative Adversarial NetworkImageMesh
🎯 What it does: This paper studies the 3D neural parametric model NeuraLeaf for leaves, which can decompose the leaf into two parts: a 2D base shape and a 3D deformation, enabling the reconstruction and generation of leaves.
NeuralSVG: An Implicit Representation for Text-to-Vector Generation
Sagi Polaczek, Daniel Cohen-Or
GenerationData SynthesisDiffusion modelScore-based ModelImageText
🎯 What it does: A method for generating text to SVG based on implicit neural networks called NeuralSVG is proposed.
Neuromanifold-Regularized KANs for Shape-fair Feature Representations
Mazlum Ferhat Arslan (Case Western Reserve University), Shuo Li (Case Western Reserve University)
ClassificationRepresentation LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a model NMR-KAN that combines Neuromanifold regularization with Kolmogorov-Arnold Networks (KAN), aiming to enhance the shape bias of visual models and suppress reliance on texture, thereby improving the model's robustness against image distortion and adversarial attacks.
Neurons: Emulating the Human Visual Cortex Improves Fidelity and Interpretability in fMRI-to-Video Reconstruction
Haonan Wang (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
RecognitionSegmentationGenerationExplainability and InterpretabilityLarge Language ModelDiffusion modelContrastive LearningVideoMultimodalityMagnetic Resonance Imaging
🎯 What it does: By simulating the hierarchical functions of the human visual cortex, the NEURONS framework is proposed, which breaks down the fMRI-to-video reconstruction task into four sub-tasks: key object segmentation, concept recognition, scene description, and fuzzy video reconstruction. The outputs of these sub-tasks are used as conditional guidance for a text-to-video diffusion model to achieve high-quality video reconstruction.
NeurOp-Diff: Continuous Remote Sensing Image Super-Resolution via Neural Operator Diffusion
Zihao Xu (Shenzhen University), Qingquan Li (Shenzhen University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A diffusion model guided by neural operators (NeurOp-Diff) is proposed to achieve continuous scale super-resolution of remote sensing images.
Neuroverse3D: Developing In-Context Learning Universal Model for Neuroimaging in 3D
Jiesi Hu (Harbin Institute of Technology), Ting Ma (Harbin Institute of Technology)
SegmentationGenerationConvolutional Neural NetworkMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Neuroverse3D is proposed, a universal model capable of handling context learning (ICL) for 3D neuroimaging;
NGD: Neural Gradient Based Deformation for Monocular Garment Reconstruction
Soham Dasgupta (Indian Institute of Technology Jodhpur), Avinash Sharma
GenerationData SynthesisNeural Radiance FieldVideoMesh
🎯 What it does: A method for clothing deformation based on neural gradient fields is proposed by recovering high-fidelity geometry and materials of dynamic clothing from monocular videos, and detail capture is achieved geometrically with adaptive remeshing.
No More Sibling Rivalry: Debiasing Human-Object Interaction Detection
Bin Yang (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: Two debiasing learning objectives are proposed to address the toxic class bias in human-object interaction detection, improving the detection transformer model.
No Pose at All: Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse Views
Ranran Huang (Imperial College London), Krystian Mikolajczyk (Imperial College London)
GenerationPose EstimationDomain AdaptationTransformerGaussian SplattingImage
🎯 What it does: This paper proposes a self-supervised, pose-independent 3D Gaussian projection framework called SPFSplat, which uses a shared ViT backbone to simultaneously predict 3D Gaussian primitives and camera poses, reconstructing sparse view scenes from uncalibrated images and performing novel view synthesis.
Noise-Modeled Diffusion Models for Low-Light Spike Image Restoration
Ruonan Liu (Beijing Institute of Technology), Hua Huang (Beijing Normal University)
RestorationSpiking Neural NetworkDiffusion modelImage
🎯 What it does: A diffusion model based on spike noise modeling has been designed and implemented for spike image restoration under low light conditions.
Noise2Score3D: Tweedie's Approach for Unsupervised Point Cloud Denoising
Xiangbin Wei (Shenzhen University), Qi Qin (City University of Hong Kong)
RestorationConvolutional Neural NetworkScore-based ModelAuto EncoderPoint Cloud
🎯 What it does: A completely unsupervised point cloud denoising framework called Noise2Score3D is proposed, which directly learns the score function of point cloud distribution from noisy data and recovers clean point clouds in one go using the Tweedie formula.
NoiseController: Towards Consistent Multi-view Video Generation via Noise Decomposition and Collaboration
Haotian Dong (Tianjin University), Qing Guo (Tianjin University)
GenerationData SynthesisAutonomous DrivingDiffusion modelVideo
🎯 What it does: This paper proposes NoiseController, a noise control framework for multi-view video generation, which enhances the spatiotemporal consistency of videos through multi-level noise decomposition, multi-frame noise collaboration, and joint denoising.
Normal and Abnormal Pathology Knowledge-Augmented Vision-Language Model for Anomaly Detection in Pathology Images
Jinsol Song (Korea University), Jin Tae Kwak (Korea University)
Anomaly DetectionTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: A framework called Ano-NAViLa is proposed, based on a visual-language model and a pool of expert terminology, for unsupervised detection of anomalies in pathological images.
NormalCrafter: Learning Temporally Consistent Normals from Video Diffusion Priors
Yanrui Bin (Hong Kong Polytechnic University), Bing Wang (Hong Kong Polytechnic University)
GenerationData SynthesisOptimizationDiffusion modelAuto EncoderVideo
🎯 What it does: We propose NormalCrafter, a surface normal estimation framework based on video diffusion models, capable of generating temporally consistent normal sequences with fine-grained details, supporting open-world videos of arbitrary lengths.
NormalLoc: Visual Localization on Textureless 3D Models using Surface Normals
Jiro Abe (NEC Corporation), Kazumine Ogura (NEC Corporation)
Pose EstimationConvolutional Neural NetworkSimultaneous Localization and MappingPoint CloudMesh
🎯 What it does: This paper proposes NormalLoc, a visual localization method that utilizes rendered normal maps as a database and training samples, capable of estimating the camera's 6-DoF pose on 3D models with missing textures and limited geometric details.
Not All Degradations Are Equal: A Targeted Feature Denoising Framework for Generalizable Image Super-Resolution
Hongjun Wang (University of Tokyo), Yinqiang Zheng (University of Tokyo)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: To address the issue of feature overfitting caused by noise in super-resolution, a pluggable target feature denoising framework is proposed, which enhances the model's generalization ability without modifying the network structure.
Not All Frame Features Are Equal: Video-to-4D Generation via Decoupling Dynamic-Static Features
Liying Yang (Macau University of Science and Technology), Yanyan Liang (Macau University of Science and Technology)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: A 4D content generation framework based on dynamic-static feature decoupling, DS4D, is proposed to address the issue of texture blurring caused by the uneven ratio of dynamic and static regions in videos.
Not all Views are Created Equal: Analyzing Viewpoint Instabilities in Vision Foundation Models
Mateusz Michalkiewicz (Rice University), Guha Balakrishnan (Rice University)
ClassificationRecognitionContrastive LearningImage
🎯 What it does: This paper systematically evaluates the feature stability of popular visual foundation models (such as CLIP, DINO, DINOv2, DreamSim, etc.) from different perspectives, proposing a method to predict perspective instability (including incidental perspectives and other perspectives) based solely on feature vectors. The significant impact of perspective instability on performance is validated in downstream tasks such as classification, VQA, and single-view 3D reconstruction.
Not Only Vision: Evolve Visual Speech Recognition via Peripheral Information
Zhaoxin Yuan (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
RecognitionTransformerLarge Language ModelMixture of ExpertsVideoTextMultimodalityAudio
🎯 What it does: This paper proposes a visual speech recognition (VSR) framework that integrates visual and peripheral information. It first uses a pre-trained visual encoder to extract lip movement features, then concatenates three types of peripheral information (contextual guidance, task expertise, and language perturbation) in the form of text prompts with visual features. Finally, it employs an improved low-rank adapter (Synergy LoRA) for autoregressive decoding on Llama3.1‑8B‑Instruct to complete the audio-to-text task.
NuiScene: Exploring Efficient Generation of Unbounded Outdoor Scenes
Han-Hung Lee (Simon Fraser University), Angel X. Chang (Simon Fraser University)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelAuto EncoderImage
🎯 What it does: This paper presents NuiScene, an efficient method for generating large-scale outdoor scenes.
NullSwap: Proactive Identity Cloaking Against Deepfake Face Swapping
Tianyi Wang (Nanyang Technological University), Yinglong Wang (Qilu University of Technology)
RecognitionSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes NullSwap, an active identity obfuscation method that embeds covert perturbations in the source image, preventing facial replacement models from extracting the original identity, thereby blocking Deepfake facial replacement in a black-box environment.
NuPlanQA: A Large-Scale Dataset and Benchmark for Multi-View Driving Scene Understanding in Multi-Modal Large Language Models
Sung-Yeon Park (Purdue University), Ziran Wang (Purdue University)
Autonomous DrivingTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the NuPlanQA-Eval evaluation benchmark and the NuPlanQA1M large-scale multi-view visual question answering dataset, and proposed the BEV-LLM baseline.
O-MaMa: Learning Object Mask Matching between Egocentric and Exocentric Views
Lorenzo Mur-Labadia (University of Zaragoza), Jose J. Guerrero (University of Zaragoza)
Object DetectionSegmentationContrastive LearningVideo
🎯 What it does: Transforming the cross-view segmentation problem into an object mask matching task, using FastSAM to generate candidate masks, and matching source masks and candidate masks through contrastive learning.
Oasis: One Image is All You Need for Multimodal Instruction Data Synthesis
Letian Zhang (Tongji University), Cheng Yang (Bytedance)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Proposes the Oasis method, which generates high-quality multimodal instruction data through MLLM autoregression using a single image and filters it through four-dimensional quality assessment.
Object-centric Video Question Answering with Visual Grounding and Referring
Haochen Wang (University of Amsterdam), Stratis Gavves (University of Amsterdam)
Object DetectionObject TrackingSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: The RGA3 model is proposed, which supports object questioning and localization in videos, capable of handling arbitrary visual prompts and generating text answers and segmentation masks.
Object-level Correlation for Few-Shot Segmentation
Chunlin Wen (Southeast University), Shuzhou Sun (Shanghai AI Laboratory)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an Object-level Correlation Network (OCNet) that achieves few-shot semantic segmentation by constructing object-level correlations between support samples and target objects in query images.
ObjectGS: Object-aware Scene Reconstruction and Scene Understanding via Gaussian Splatting
Ruijie Zhu (University of Science and Technology of China), Bo Dai (University of Hong Kong)
Object DetectionSegmentationGaussian SplattingPoint Cloud
🎯 What it does: This paper presents ObjectGS, an object-aware framework based on Gaussian Splatting, achieving a unification of 3D scene reconstruction and semantic understanding.
ObjectMate: A Recurrence Prior for Object Insertion and Subject-Driven Generation
Daniel Winter (Google), Yedid Hoshen (Hebrew University of Jerusalem)
Object DetectionGenerationData SynthesisDiffusion modelImage
🎯 What it does: A zero-shot tuning method named ObjectMate is proposed, which utilizes the prior of object repetition to synthesize multi-view reference objects with scene descriptions into images, achieving high-fidelity identity preservation and lighting matching.
ObjectRelator: Enabling Cross-View Object Relation Understanding Across Ego-Centric and Exo-Centric Perspectives
Yuqian Fu (INSAIT), Luc Van Gool (INSAIT)
Object DetectionSegmentationTransformerLarge Language ModelImageVideoMultimodality
🎯 What it does: The ObjectRelator method is proposed to address the object correspondence segmentation task between self-perspective and external perspective, achieving cross-perspective consistency through multimodal conditional fusion and self-supervised cross-perspective alignment.
OccluGaussian: Occlusion-Aware Gaussian Splatting for Large Scene Reconstruction and Rendering
Shiyong Liu (Huawei Noah's Ark Lab), Xiaofei Wu (Huawei Noah's Ark Lab)
RestorationGenerationGraph Neural NetworkGaussian SplattingPoint Cloud
🎯 What it does: Proposes an Occlusion-Aware 3D Gaussian splatting method for high-quality reconstruction and real-time rendering of large-scale scenes;
Occlusion-robust Stylization for Drawing-based 3D Animation
Sunjae Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
Image TranslationGenerationDiffusion modelContrastive LearningOptical FlowImage
🎯 What it does: A robust style transfer framework based on flow depth edge detection (OSF) is proposed to address the style degradation problem caused by pose differences in hand-drawn 3D animations.
Occupancy Learning with Spatiotemporal Memory
Ziyang Leng (University of California), Bolei Zhou (University of California)
SegmentationAutonomous DrivingConvolutional Neural NetworkSupervised Fine-TuningPoint CloudBenchmark
🎯 What it does: This paper studies a 3D occupancy prediction framework ST-Occ based on scene-level spatiotemporal memory.
OCK: Unsupervised Dynamic Video Prediction with Object-Centric Kinematics
Yeon-Ji Song (Seoul National University), Byoung-Tak Zhang (Seoul National University)
GenerationAutonomous DrivingTransformerVideo
🎯 What it does: Proposed an object-centered motion prediction model OCK, which utilizes object slots and explicit kinematic features to achieve unsupervised video prediction.
OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation
Junyuan Zhang (Shanghai AI Laboratory), Wentao Zhang (Peking University)
GenerationRetrievalTransformerVision Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed OHRBench to evaluate the impact of OCR noise on Retrieval-Augmented Generation (RAG) systems.
OcRFDet: Object-Centric Radiance Fields for Multi-View 3D Object Detection in Autonomous Driving
Mingqian Ji (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Object DetectionAutonomous DrivingNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes OcRFDet, which enhances the geometric feature representation and foreground modeling of multi-view 3D object detection through a dual-module approach of Object-Centered Radiance Field (OcRF) and Height-Aware Transparency Attention (HOA).
OCSplats: Observation Completeness Quantification and Label Noise Separation in 3DGS
Han Ling (Nanjing University of Science and Technology), Quansen Sun (Southeast University)
RestorationGaussian SplattingPoint Cloud
🎯 What it does: Conduct noise-robust reconstruction for 3D Gaussian Splatting (3DGS) and propose the Observation Completeness (OC) metric to assess the observational sufficiency of each region. Based on OC, perform noise assessment correction and efficient pruning, ultimately achieving self-supervised separation of label noise and high-quality 3D scene reconstruction.
OD-RASE: Ontology-Driven Risk Assessment and Safety Enhancement for Autonomous Driving
Kota Shimomura (Chubu University), Takayuki Kawabuchi (Honda R and D Company Limited)
Autonomous DrivingSafty and PrivacyTransformerLarge Language ModelVision Language ModelDiffusion modelImageMultimodalityChain-of-Thought
🎯 What it does: The OD-RASE framework is proposed, which utilizes expert knowledge to construct an ontology for filtering road structures and accident risks through graph matching, and automatically generates renovation suggestions and visualizations using large-scale visual language models and diffusion models, thereby building a high-quality multimodal dataset.
ODDR: Outlier Detection & Dimension Reduction Based Defense Against Adversarial Patches
Nandish Chattopadhyay (New York University), Muhammad Shafique (Technology Innovation Institute)
Object DetectionDepth EstimationAnomaly DetectionImage
🎯 What it does: A three-stage model-agnostic defense framework called ODDR is proposed, which utilizes chunking and Isolation Forest anomaly detection to locate adversarial patches, and neutralizes the detected patch areas through SVD dimensionality reduction, thereby resisting patch-based attacks while maintaining the original model performance.
ODP-Bench: Benchmarking Out-of-Distribution Performance Prediction
Han Yu (Tsinghua University), Peng Cui (Renmin University of China)
Domain AdaptationAnomaly DetectionConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImageBenchmark
🎯 What it does: A comprehensive OOD performance prediction benchmark, ODP-Bench, is proposed, covering 29 different types of OOD datasets, 10 mainstream performance prediction algorithms, and publicly releasing 1444 pre-trained models to form a unified evaluation environment.
Omegance: A Single Parameter for Various Granularities in Diffusion-Based Synthesis
Xinyu Hou (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
GenerationData SynthesisDiffusion modelImageVideoText
🎯 What it does: The Omegance method is proposed, which controls the detail granularity of generated images through a noise scaling parameter ω during the reverse sampling process of diffusion models, enabling detail control on global, spatial (mask), and temporal (scheduling) levels;
OminiControl: Minimal and Universal Control for Diffusion Transformer
Zhenxiong Tan (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageText
🎯 What it does: Design and implement OminiControl, a minimalist image control framework that utilizes the existing VAE encoder and Transformer blocks of the Diffusion Transformer, supporting various control tasks with spatial alignment and non-alignment.
OMNI-DC: Highly Robust Depth Completion with Multiresolution Depth Integration
Yiming Zuo (Princeton University), Jia Deng (Princeton University)
RestorationDepth EstimationTransformerImage
🎯 What it does: A zero-shot deep completion model OMNI-DC is proposed, which can directly generate high-quality dense depth maps under different datasets and sparse depth patterns.
Omni-scene Perception-oriented Point Cloud Geometry Enhancement for Coordinate Quantization
Wang Liu (Peking University), Wei Gao (Peking University)
Object DetectionCompressionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This study investigates the issue of point cloud coordinate quantization distortion and proposes a unified root-growing-pruning mechanism to enhance the geometric quality of both small-scale and large-scale point clouds.
OmniCache: A Trajectory-Oriented Global Perspective on Training-Free Cache Reuse for Diffusion Transformer Models
Huanpeng Chu (Zhipu AI), Yutao Zhang (Zhipu AI)
GenerationOptimizationComputational EfficiencyTransformerDiffusion modelImageVideoText
🎯 What it does: Training acceleration for diffusion Transformers is free.
OmniDiff: A Comprehensive Benchmark for Fine-grained Image Difference Captioning
Yuan Liu (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A fine-grained image difference description dataset, OmniDiff, covering 324 complex real and 3D synthetic scenes has been constructed, and a pluggable Multi-scale Differential Perception (MDP) module has been proposed based on a multi-modal large language model (MLLM), forming the M3Diff model to generate detailed and semantically complete difference descriptions.
OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models
Gaojie Lin (ByteDance Intelligent Creation), Jingtuo Liu (ByteDance Intelligent Creation)
GenerationData SynthesisTransformerDiffusion modelVideoTextMultimodalityAudio
🎯 What it does: Developed OmniHuman, an end-to-end multimodal conditional human video generation framework that can animate characters based solely on an image and audio, pose, or text.
OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting
Yongsheng Yu (University of Rochester), Jiebo Luo (University of Rochester)
Image TranslationRestorationGenerationTransformerDiffusion modelImage
🎯 What it does: A unified framework called OmniPaint has been developed, enabling interdependent image editing for object removal and insertion.