ICCV 2025 Papers — Page 16
IEEE/CVF International Conference on Computer Vision · 2701 papers
MeshAnything V2: Artist-Created Mesh Generation with Adjacent Mesh Tokenization
Yiwen Chen (Nanyang Technological University), Guosheng Lin (Nanyang Technological University)
GenerationData SynthesisTransformerPoint CloudMesh
🎯 What it does: Designed and implemented MeshAnything V2, a model based on autoregressive Transformer, to generate artist-created triangle meshes aligned with given shapes, and proposed the Adjacent Mesh Tokenization (AMT) method.
MeshLLM: Empowering Large Language Models to Progressively Understand and Generate 3D Mesh
Shuangkang Fang (Beihang University), Ming-Hsuan Yang (University of California Merced)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPoint CloudMesh
🎯 What it does: This paper proposes the MeshLLM framework, enabling large language models (LLMs) to directly understand and generate text-serialized 3D meshes.
MeshMamba: State Space Models for Articulated 3D Mesh Generation and Reconstruction
Yusuke Yoshiyasu (National Institute of Advanced Industrial Science and Technology), Ryusuke Sagawa (National Institute of Advanced Industrial Science and Technology)
GenerationData SynthesisPose EstimationTransformerDiffusion modelMesh
🎯 What it does: Proposes MeshMamba to learn and generate dense 3D joint meshes using Mamba-SSMs, achieving complete reconstruction and generation of the human body (including hands, face, and clothing).
MeshPad: Interactive Sketch-Conditioned Artist-Reminiscent Mesh Generation and Editing
Haoxuan Li (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationTransformerMesh
🎯 What it does: This paper presents MeshPad, a system for interactively generating and editing triangular meshes through sketching, capable of performing addition and deletion operations on existing meshes until complex models are completed.
Met2Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems
Shaohan Li (Chengdu University of Information Technology), Xiaolin Qin (Chengdu Institute of Computer Applications, Chinese Academy of Sciences)
OptimizationTransformerAuto EncoderTime Series
🎯 What it does: For multivariate weather forecasting, an implicit two-stage training Met 2Net model is proposed, which uses independent encoders/decoders and a self-attention translator in the latent space to achieve interaction and prediction among variables.
Meta-Learning Dynamic Center Distance: Hard Sample Mining for Learning with Noisy Labels
Chenyu Mu (Xidian University), Cheng Deng (Xidian University)
ClassificationSegmentationMeta LearningConvolutional Neural NetworkContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: For learning tasks with noisy labels, a hard sample mining method based on Dynamic Center Distance (DCD) is proposed. This method dynamically updates class centers and sample weights through a meta-learning framework, combining pseudo-labels and semi-supervised learning to enhance the model's robustness on noisy data.
Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Concepts
Hongcheng Gao (Sea AI Lab), Min Lin (Sea AI Lab)
GenerationMeta LearningDiffusion modelImage
🎯 What it does: A meta-learning based 'Meta-Unlearning' framework is proposed, which combines existing de-conceptualization methods to prevent diffusion models from relearning forgotten concepts when maliciously fine-tuned, and causes related retained concepts to self-destruct.
MetaMorph: Multimodal Understanding and Generation via Instruction Tuning
Shengbang Tong (Meta), Zhuang Liu (Meta)
RecognitionGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Through Visual Prediction Instruction Tuning (VPiT), the pre-trained LLM is transformed into a unified multimodal model MetaMorph, which can understand images and generate visual tokens.
MetaScope: Optics-Driven Neural Network for Ultra-Micro Metalens Endoscopy
Wuyang Li (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
RestorationSegmentationKnowledge DistillationConvolutional Neural NetworkMixture of ExpertsImageBiomedical Data
🎯 What it does: This paper proposes MetaScope, an optical-driven neural network for images from metal lens endoscopes, achieving joint segmentation and distortion correction.
METEOR: Multi-Encoder Collaborative Token Pruning for Efficient Vision Language Models
Yuchen Liu (Shanghai Jiao Tong University), Qi Tian (Huawei Inc.)
RecognitionComputational EfficiencyTransformerVision Language ModelImageTextMultimodality
🎯 What it does: The METEOR framework for multi-encoder collaborative token pruning is proposed, which can gradually remove redundant visual tokens in the three stages of visual encoding, fusion, and LLM decoding.
Metric Convolutions: A Unifying Theory to Adaptive Image Convolutions
Thomas Dagès (Technion - Israel Institute of Technology), Alfred M. Bruckstein (Technion - Israel Institute of Technology)
ClassificationRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes Metric Convolution based on image manifolds, which achieves adaptive convolution kernels by sampling unit spheres (tangent spaces or geodesics) at each pixel, unifying and extending existing variable convolution methods.
MGSfM: Multi-Camera Geometry Driven Global Structure-from-Motion
Peilin Tao (Institute of Automation, Chinese Academy of Sciences), Shuhan Shen (Institute of Automation, Chinese Academy of Sciences)
Pose EstimationAutonomous DrivingOptimizationSimultaneous Localization and MappingImageVideo
🎯 What it does: A multi-camera driven global structure from motion (SfM) framework called MGSfM is proposed, which combines decoupled rotation averaging and mixed translation averaging to achieve efficient and robust global camera pose estimation and 3D reconstruction.
MGSR: 2D/3D Mutual-boosted Gaussian Splatting for High-fidelity Surface Reconstruction under Various Light Conditions
Qingyuan Zhou (Fudan University), Ying He (Nanyang Technological University)
RestorationGenerationGaussian SplattingPoint CloudMesh
🎯 What it does: A dual-branch mutually reinforcing 2D/3D Gaussian scattering framework MGSR is proposed for achieving high-fidelity surface reconstruction and novel view synthesis under varying lighting conditions.
MH-LVC: Multi-Hypothesis Temporal Prediction for Learned Conditional Residual Video Coding
Huu-Tai Phung (National Yang Ming Chiao Tung University), Wen-Hsiao Peng (National Yang Ming Chiao Tung University)
CompressionConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: A multi-hypothesis temporal prediction scheme (MH-LVC) is proposed in learning-based video compression, which constructs a temporal predictor by simultaneously using short-term recent frames and long-term key frames, and utilizes the prediction results for conditional residual coding.
MiDSummer: Multi-Guidance Diffusion for Controllable Zero-Shot Immersive Gaussian Splatting Scene Generation
Anjun Hu (Amazon Prime Video), Hanting Xie (Amazon Prime Video)
GenerationData SynthesisLarge Language ModelDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: MiDSummer achieves zero-shot controllable immersive Gaussian point cloud scene generation through a two-stage multi-guided diffusion framework, first using LLM and GDM for parallel layout planning, and then refining from multiple perspectives guided by the layout.
MIEB: Massive Image Embedding Benchmark
Chenghao Xiao (Durham University), Niklas Muennighoff (Contextual AI)
RetrievalRepresentation LearningContrastive LearningImageMultimodalityBenchmark
🎯 What it does: A large-scale image embedding benchmark MIEB is proposed, covering 130 tasks, 38 languages, and evaluating 50 models.
MikuDance: Animating Character Art with Mixed Motion Dynamics
Jiaxu Zhang (Wuhan University), Zhigang Tu (Wuhan University)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: MikuDance is proposed, a pipeline based on diffusion models that can generate high-quality animations from character art images using high dynamics and mixed motion instructions.
MinCD-PnP: Learning 2D-3D Correspondences with Approximate Blind PnP
Pei An (Huazhong University of Science and Technology), Liangliang Nan (Huazhong University of Science and Technology)
Pose EstimationConvolutional Neural NetworkSupervised Fine-TuningImagePoint Cloud
🎯 What it does: This paper proposes a new image-point cloud registration framework MinCD-PnP and MinCD-Net, which utilizes the minimization of the Chamfer distance approximating blind PnP to learn the 2D-3D correspondence, thereby achieving high-precision camera pose estimation.
Mind the Cost of Scaffold! Benign Clients May Even Become Accomplices of Backdoor Attack
Xingshuo Han (Nanjing University of Aeronautics and Astronautics), Tianwei Zhang (Nanyang Technological University)
Federated LearningAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes and implements a persistent backdoor attack called BadSFL for Scaffold Federated Learning, utilizing control variables and GAN for data augmentation, allowing ordinary clients to unknowingly assist in propagating the backdoor.
Mind the Gap: Aligning Vision Foundation Models to Image Feature Matching
Yuhan Liu (Xi'an Jiaotong University), Jingmin Xin (Xi'an Jiaotong University)
RetrievalDiffusion modelImageBenchmark
🎯 What it does: This paper proposes the IMD framework, which extracts image features through a generative diffusion model and achieves instance-level matching using a cross-image interaction prompt module.
Mind the Gap: Preserving and Compensating for the Modality Gap in CLIP-Based Continual Learning
Linlan Huang (Nankai University), Xialei Liu (Nankai University)
ClassificationRepresentation LearningTransformerContrastive LearningImageMultimodality
🎯 What it does: Proposes a replay-free class-incremental learning approach by preserving and compensating for the modal gap of the CLIP model.
MINERVA: Evaluating Complex Video Reasoning
Arsha Nagrani (Google DeepMind), Tobias Weyand (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringVideoMultimodality
🎯 What it does: The MINERVA dataset was constructed, containing 1515 questions from multiple domains and varying video lengths, each with 5 answer options and manually crafted multi-step reasoning trajectories to evaluate the complex video reasoning capabilities of multimodal models.
MIORe & VAR-MIORe: Benchmarks to Push the Boundaries of Restoration
George Ciubotariu (University of Wurzburg), Radu Timofte (University of Wurzburg)
RestorationOptical FlowVideoBenchmark
🎯 What it does: Two multi-task datasets, MIORe and VAR-MIORe, are proposed, covering motion recovery tasks from low to extreme motion amplitudes, and algorithms such as deblurring, frame interpolation, and optical flow are evaluated on them.
MissRAG: Addressing the Missing Modality Challenge in Multimodal Large Language Models
Vittorio Pipoli (University of Modena and Reggio Emilia), Elisa Ficarra (University of Modena and Reggio Emilia)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: This study investigates and evaluates the robustness of multimodal large language models (MLLM) in the absence of modalities and proposes a missing modality compensation framework called MISSRAG, which combines retrieval-augmented generation (RAG) with prompt engineering.
MistSense: Versatile Online Detection of Procedural and Execution Mistakes
Constantin Patsch (Technical University of Munich), Eckehard Steinbach (Technical University of Munich)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: An online error detection and explanation system called MistSense is proposed and implemented, capable of identifying two types of errors: procedural and execution errors, and generating explanatory text using a large language model upon detecting an error.
Mitigating Catastrophic Overfitting in Fast Adversarial Training via Label Information Elimination
Chao Pan (Southern University of Science and Technology), Xin Yao (Lingnan University)
OptimizationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper reveals the phenomenon of Catastrophic Overfitting (CO) in Fast Adversarial Training by studying the transferability of label information across samples in single-step adversarial perturbations, and proposes the Label Information Elimination Training (LIET) method to actively eliminate such label information, thereby reducing the risk of CO.
Mitigating Geometric Degradation in Fast DownSampling via FastAdapter for Point Cloud Segmentation
Shuofeng Sun (Beijing University of Posts and Telecommunications), Haibin Yan (Beijing University of Posts and Telecommunications)
SegmentationComputational EfficiencyKnowledge DistillationPoint Cloud
🎯 What it does: Proposes the FastAdapter, which aggregates local information using anchor points and backpropagates it to downsampled points, addressing the issue of geometric information degradation caused by random sampling.
Mitigating Object Hallucinations via Sentence-Level Early Intervention
Shangpin Peng (Harbin Institute of Technology), Zhuotao Tian (Harbin Institute of Technology)
Object DetectionLarge Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the SENTINEL framework, which addresses the object hallucination problem in multimodal large language models by employing sentence-level early intervention to suppress the spread of hallucinations.
MixA-Q: Revisiting Activation Sparsity for Vision Transformers from a Mixed-Precision Quantization Perspective
Weitian Wang (Robert Bosch GmbH), Akash Kumar (Ruhr University Bochum)
Object DetectionSegmentationComputational EfficiencyKnowledge DistillationTransformerImage
🎯 What it does: This paper proposes MixA-Q, a mixed-precision activation quantization framework designed for window-based visual Transformers, which dynamically allocates high and low precision based on window-level activation sparsity.
MixA: A Mixed Attention approach with Stable Lightweight Linear Attention to enhance Efficiency of Vision Transformers at the Edge
Sabbir Ahmed (Binghamton University), Lingjuan Lyu (Sony AI)
ClassificationObject DetectionSegmentationComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningImage
🎯 What it does: A hybrid attention framework called MixA is designed, which modifies some layers of the Vision Transformer to use ReLU-based quadratic attention, while other layers are changed to the linear attention module SteLLA, in order to improve inference speed while maintaining performance.
MixANT: Observation-dependent Memory Propagation for Stochastic Dense Action Anticipation
Syed Talal Wasim (University of Bonn), Juergen Gall (University of Bonn)
Mixture of ExpertsDiffusion modelVideoSequential
🎯 What it does: The MixANT architecture is proposed, which combines a mixture of experts mechanism to dynamically select the forgetting matrix A in the Mamba model, achieving input-dependent memory propagation and generating diverse future action predictions in random long sequence action forecasting.
Mixed Signals: A Diverse Point Cloud Dataset for Heterogeneous LiDAR V2X Collaboration
Katie Z Luo (Cornell University), Julie Stephany Berrio Perez (University of Sydney)
Object DetectionAutonomous DrivingSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: Created and released the Mixed Signals V2X dataset, which includes 45.1k point clouds and 240.6k bounding boxes, covering 10 fine-grained categories, along with real left-driving scenarios involving 3 vehicles (different LiDAR poses) and a dual LiDAR RSU; benchmarked various collaborative perception methods (early/mid/late/Laly cascade) on this dataset.
MixRI: Mixing Features of Reference Images for Novel Object Pose Estimation
Xinhang Liu (Northwestern Polytechnical University), Yuchao Dai (Northwestern Polytechnical University)
Pose EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a lightweight network MixRI, which achieves 6DoF pose estimation for unseen objects in RGB images by directly matching a small number of reference images, without the need for retrieval and pre-cached features.
Mixture of Experts Guided by Gaussian Splatters Matters: A new Approach to Weakly-Supervised Video Anomaly Detection
Giacomo D' Amicantonio, Egor Bondarev (Eindhoven University of Technology)
Anomaly DetectionTransformerMixture of ExpertsGaussian SplattingVideoMultimodality
🎯 What it does: A multi-expert (Mixture-of-Experts) framework based on Gaussian splatting is proposed, utilizing Temporal Gaussian Splatting loss to enhance the temporal integrity of weakly supervised video anomaly detection, and integrating experts with coarse-grained features through a gating model.
Mixture-of-Scores: Robust Image-Text Data Valuation via Three Lines of Code
Sitong Wu (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
Data-Centric LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper studies the differences between image-text pairing quality scoring models and proposes the Mixture-of-Scores (MoS) method, which integrates multiple scoring models to achieve a more robust quality assessment.
MM-IFEngine: Towards Multimodal Instruction Following
Shengyuan Ding (Fudan University), Jiaqi Wang (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityBenchmark
🎯 What it does: Proposed the MM-IFEngine pipeline, which automatically generates image-instruction pairs with various constraints, and based on this, constructs MM-IFInstruct-23k (SFT data), MM-IFDPO-23k (DPO data), and the MM-IFEval evaluation benchmark;
MM-Spatial: Exploring 3D Spatial Understanding in Multimodal LLMs
Erik Daxberger (Apple), Peter Grasch (Apple)
RecognitionGenerationDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityPoint Cloud
🎯 What it does: This paper proposes a multimodal large language model, MM-Spatial, for understanding three-dimensional space, and creates a new VQA dataset, CA-VQA, through a generation pipeline based on high-quality 3D scene data. The model is then fine-tuned with supervision on this dataset and evaluated on multiple 3D spatial tasks.
MMAD: Multi-label Micro-Action Detection in Videos
Kun Li (Hefei University of Technology), Meng Wang (Zhejiang University)
RecognitionObject DetectionTransformerVideo
🎯 What it does: This paper proposes the task of Multi-Label Micro Action Detection (MMAD) and constructs the MMA-52 dataset, providing a baseline model and conducting experimental evaluations.
MMAIF: Multi-task and Multi-degradation All-in-One for Image Fusion with Language Guidance
Zihan Cao (University of Electronic Science and Technology of China), Liang-Jian Deng (University of Electronic Science and Technology of China)
RestorationGenerationTransformerLarge Language ModelMixture of ExpertsDiffusion modelImageText
🎯 What it does: A unified multi-task, multi-degradation, language-guided image fusion framework (MMAIF) is proposed, which can simultaneously perform image degradation processing such as denoising, deraining, and defogging, as well as multi-modal/multi-exposure/multi-focus fusion within the same model.
MMAT-1M: A Large Reasoning Dataset for Multimodal Agent Tuning
Tianhong Gao (Baidu Inc), Gang Zhang (Baidu Inc)
OptimizationData-Centric LearningTransformerLarge Language ModelTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: We constructed and released MMAT-1M, the first multimodal agent tuning dataset at a million scale, which includes reasoning steps, tool calls, and reflection processes generated based on GPT-4o, and provides both one-round (ORR) and multi-round (RR) formats; at the same time, we evaluated the model's performance improvement on eight major public benchmarks and Dyn-VQA.
mmCooper: A Multi-agent Multi-stage Communication-efficient and Collaboration-robust Cooperative Perception Framework
Bingyi Liu (Wuhan University Of Technology), Libing Wu (Wuhan University Of Technology)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: The mmCooper framework is proposed for multi-vehicle collaborative perception, achieving multi-stage information fusion and communication savings.
MMCR: Benchmarking Cross-Source Reasoning in Scientific Papers
Yang Tian (Shanghai Jiao Tong University), Bo Zhao (Shanghai Jiao Tong University)
Vision Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A benchmark called MMCR was constructed, with 276 manually annotated cross-source reasoning questions to evaluate the cross-source reasoning ability of VLM in scientific papers.
MMGeo: Multimodal Compositional Geo-Localization for UAVs
Yuxiang Ji (Xiamen University), Liaoni Wu (Xiamen University)
RetrievalAutonomous DrivingTransformerContrastive LearningImageTextMultimodalityPoint Cloud
🎯 What it does: This paper proposes MMGEO, which utilizes multi-modal combined queries (image + point cloud/depth/text) to achieve UAV geographic positioning in satellite image retrieval databases.
MMOne: Representing Multiple Modalities in One Scene
Zhifeng Gu (Hong Kong Polytechnic University), Bing Wang (Hong Kong Polytechnic University)
Representation LearningNeural Radiance FieldGaussian SplattingImageTextMultimodality
🎯 What it does: Proposes the MMOne framework, which integrates multimodal scene representation and addresses attribute and granularity conflicts.
MMReason: An Open-Ended Multi-Modal Multi-Step Reasoning Benchmark for MLLMs Toward AGI
Huanjin Yao (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
Large Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A benchmark called MMReason is proposed and constructed to evaluate the long-chain reasoning ability of multimodal LLMs;
Mobile Video Diffusion
Haitam Ben Yahia (Qualcomm AI Research), Amirhossein Habibian (Qualcomm AI Research)
GenerationData SynthesisComputational EfficiencyDiffusion modelVideo
🎯 What it does: An image-to-video diffusion model called MobileVD has been implemented on mobile devices, capable of quickly generating 14 frames of 512×256 resolution video in low-power environments.
MobileIE: An Extremely Lightweight and Effective ConvNet for Real-Time Image Enhancement on Mobile Devices
Hailong Yan (University of Electronic Science and Technology of China), Le Zhang (Hefei University of Technology)
RestorationOptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A highly lightweight CNN MobileIE with only 4K parameters is proposed for real-time image enhancement on mobile devices.
MobileViCLIP: An Efficient Video-Text Model for Mobile Devices
Min Yang (Nanjing University), Limin Wang (Nanjing University)
RecognitionRetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A lightweight video-text model called MobileViCLIP has been developed for mobile devices. By adding spatiotemporal reparameterization modules (Spatiotemporal RepMixer and Spatiotemporal Attention) to the existing efficient image-text model MobileCLIP and fine-tuning it on a large-scale video-text dataset, the model achieves high-speed inference directly on mobile phones while possessing strong zero-shot retrieval and action recognition capabilities.
MOBIUS: Big-to-Mobile Universal Instance Segmentation via Multi-modal Bottleneck Fusion and Calibrated Decoder Pruning
Mattia Segu (Google), Federico Tombari (Google)
Object DetectionSegmentationTransformerVision Language ModelImageMultimodality
🎯 What it does: Developed the MOBIUS series models to achieve general instance segmentation from large-scale to mobile devices;
ModalTune: Fine-Tuning Slide-Level Foundation Models with Multi-Modal Information for Multi-task Learning in Digital Pathology
Vishwesh Ramanathan (Sunnybrook Research Institute), Anne L. Martel (Sunnybrook Research Institute)
ClassificationSegmentationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBiomedical Data
🎯 What it does: The ModalTune framework is proposed, utilizing SLFM (such as Gigapath) to achieve multi-modal, multi-task, and pan-cancer joint fine-tuning in digital pathology, significantly improving subtype prediction and survival prediction performance.
Model Reveals What to Cache: Profiling-Based Feature Reuse for Video Diffusion Models
Xuran Ma (Hong Kong University of Science and Technology), Harry Yang (Hong Kong University of Science and Technology)
GenerationComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: For the video diffusion model DiT, we propose the ProfilingDiT mechanism, which achieves adaptive feature caching based on attention and noise variations during forward inference, significantly reducing inference costs without sacrificing visual quality.
Modeling Human Gaze Behavior with Diffusion Models for Unified Scanpath Prediction
Giuseppe Cartella (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
TransformerDiffusion modelImage
🎯 What it does: ScanDiff is proposed to generate diverse, task-driven visual scanning paths through diffusion models and Vision Transformers.
Modeling Saliency Dataset Bias
Matthias Kümmerer (Tübingen AI Center University of Tübingen), Matthias Bethge (Tübingen AI Center University of Tübingen)
Domain AdaptationExplainability and InterpretabilityConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The study investigates the bias in visual attention datasets and proposes a saliency model that requires only 20 interpretable parameters for cross-dataset adaptation.
Moderating the Generalization of Score-based Generative Model
Wan Jiang (Hefei University of Technology), Richang Hong (Hefei University of Technology)
GenerationData SynthesisDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: A machine forgetting framework (MSGM) for Score-based Generative Models is proposed, achieving active suppression of non-sought-for generated data (NSFG) while maintaining the quality of sought-for generated data (SFG).
ModSkill: Physical Character Skill Modularization
Yiming Huang (University of Pennsylvania), Lingjie Liu (University of Pennsylvania)
Robotic IntelligenceReinforcement LearningDiffusion modelVideoText
🎯 What it does: The ModSkill framework is proposed, which decomposes full-body movements into reusable skills at the body part level and achieves high-precision motion tracking through an attention mechanism and low-level controllers.
MOERL: When Mixture-of-Experts Meet Reinforcement Learning for Adverse Weather Image Restoration
Tao Wang (Nanjing University), Wenhan Luo (Hong Kong University of Science and Technology)
RestorationTransformerReinforcement LearningMixture of ExpertsVision Language ModelImage
🎯 What it does: A unified weather image restoration framework MOERL that combines Mixture-of-Experts (MoE) and reinforcement learning is proposed;
MoFRR: Mixture of Diffusion Models for Face Retouching Restoration
Jiaxin Liu (Fudan University), Xinpeng Zhang (Fudan University)
RestorationMixture of ExpertsDiffusion modelImage
🎯 What it does: This paper proposes the Facial Retouching Recovery (FRR) task and designs MoFRR—a mixed expert diffusion model framework to recover original portraits from beautified facial images.
MoGA: 3D Generative Avatar Prior for Monocular Gaussian Avatar Reconstruction
Zijian Dong, Andreas Geiger
GenerationData SynthesisPose EstimationDiffusion modelImage
🎯 What it does: By combining a 3D generative avatar prior with a multi-view diffusion model, a high-fidelity 3D Gaussian face animation model is reconstructed from a single image.
MolParser: End-to-end Visual Recognition of Molecule Structures in the Wild
Xi Fang (DP Technology), Guolin Ke (DP Technology)
RecognitionObject DetectionGenerationData SynthesisTransformerSupervised Fine-TuningImageText
🎯 What it does: An end-to-end Optical Chemical Structure Recognition (OCSR) framework called MolParser is proposed, which can directly convert chemical structure images into textual chemical descriptions. A dataset containing 7.7M aligned image-E-SMILES pairs, named MolParser-7M, was constructed, and approximately 400k real samples were collected and manually annotated from actual patents and papers using an active learning approach.
MoMa-Kitchen: A 100K+ Benchmark for Affordance-Grounded Last-Mile Navigation in Mobile Manipulation
Pingrui Zhang (Fudan University), Xuelong Li (TeleAI)
Robotic IntelligenceSimultaneous Localization and MappingMultimodalityPoint CloudBenchmark
🎯 What it does: A large-scale kitchen scene dataset, MoMa-Kitchen (approximately 127k robot navigation-manipulation records), was constructed, and the NavAff model was proposed to achieve feasibility prediction for navigation based on vision and point clouds, supporting robots in locating the optimal chassis position for executable grasping in complex environments.
MoMaps: Semantics-Aware Scene Motion Generation with Motion Maps
Jiahui Lei (University of Pennsylvania), Leonidas Guibas (Google DeepMind)
Object TrackingGenerationData SynthesisDepth EstimationVision Language ModelDiffusion modelGaussian SplattingVideoMultimodality
🎯 What it does: This paper proposes a pixel-aligned three-dimensional motion representation called Motion Map (MoMap) and constructs a MoMap database using a large-scale real video dataset (approximately 50,000 videos). It trains a diffusion model based on Stable Diffusion to generate future three-dimensional dense motion from single-frame RGB images and optional text. The generated MoMap is then rendered into partial videos, and a lightweight image diffusion model is used to complete the missing areas, achieving two-dimensional video synthesis. Additionally, a structured DSL generated by a Vision-Language model is introduced to enhance motion control accuracy.
Moment Quantization for Video Temporal Grounding
Xiaolong Sun (Xi'an Jiaotong University), Gang Hua (Amazon)
RecognitionRetrievalTransformerContrastive LearningVideo
🎯 What it does: This paper proposes the Moment Quantization (MQVTG) method, which maps continuous video features to discrete codewords through vector quantization, enhancing the discriminative ability between relevant and irrelevant segments in video temporal localization.
Momentum-GS: Momentum Gaussian Self-Distillation for High-Quality Large Scene Reconstruction
Jixuan Fan (Tsinghua University), Yansong Tang (Tsinghua University)
Knowledge DistillationRepresentation LearningGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes Momentum-GS, which enhances the stability and consistency of 3D Gaussian Splatting in large-scale scene reconstruction through momentum self-distillation.
Monocular Facial Appearance Capture in the Wild
Yingyan Xu (ETH Zurich), Derek Bradley (Disney Research Studios)
GenerationPose EstimationOptimizationNeural Radiance FieldImageVideo
🎯 What it does: This paper proposes a lightweight method that captures facial geometry and appearance (diffuse, specular intensity, and roughness) in outdoor environments using only a single camera head rotation video, capable of generating high-quality re-lightable 3D facial assets under any lighting conditions.
Monocular Semantic Scene Completion via Masked Recurrent Networks
Xuzhi Wang (Tianjin Normal University), Ziping Zhao (Tianjin Normal University)
SegmentationDepth EstimationAutonomous DrivingRecurrent Neural NetworkImagePoint Cloud
🎯 What it does: A two-stage monocular semantic scene completion framework MonoMRN is proposed, which first makes a rough estimate and then iteratively refines it through a mask recursive network.
MonoFusion: Sparse-View 4D Reconstruction via Monocular Fusion
Zihan Wang (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
GenerationDepth EstimationGaussian SplattingVideoPoint Cloud
🎯 What it does: The research conducts 4D reconstruction of dynamic scenes from a sparse perspective (four fixed cameras) and proposes the MonoFusion framework, which can reconstruct dynamic human behaviors from a limited number of viewpoints and generate high-quality renderings from new perspectives.
MonoMobility: Zero-Shot 3D Mobility Analysis from Monocular Videos
Hongyi Zhou (National University of Defense Technology), Kai Xu (National University of Defense Technology)
Pose EstimationDepth EstimationOptimizationOptical FlowVideoPoint Cloud
🎯 What it does: A zero-shot framework called MonoMobility is proposed to analyze the motion components and motion attributes of articulated objects from monocular video.
MonoMVSNet: Monocular Priors Guided Multi-View Stereo Network
Jianfei Jiang (University of Science and Technology Beijing), Huimin Ma (University of Science and Technology Beijing)
Depth EstimationTransformerPoint Cloud
🎯 What it does: This paper proposes MonoMVSNet, a multi-view stereo network guided by monocular priors, which enhances multi-view matching and depth estimation using features and depth information from a monocular depth model.
MonoSOWA: Scalable Monocular 3D Object Detector Without Human Annotations
Jan Skvrna (Czech Technical University in Prague), Lukas Neumann (Czech Technical University in Prague)
Object DetectionDepth EstimationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Proposes an automatic annotation and training framework for training 3D object detection models from a monocular camera without manual labeling and without LiDAR.
monoVLN: Bridging the Observation Gap between Monocular and Panoramic Vision and Language Navigation
Renjie Lu (Sun Yat-sen University), Wei-Shi Zheng
Robotic IntelligenceContrastive LearningGaussian SplattingImage
🎯 What it does: Design the monoVLN framework to address the issue of incomplete visual information under a monocular RGB-D camera, and propose implicit partial completion and uncertainty-based active perception strategies;
MonSTeR: a Unified Model for Motion, Scene, Text Retrieval
Luca Collorone (Sapienza University of Rome), Fabio Galasso (Sapienza University of Rome)
RetrievalTransformerAuto EncoderContrastive LearningTextMultimodality
🎯 What it does: The MonSTeR model is proposed to achieve unified retrieval and evaluation of text, actions, and scenes in a tri-modal manner.
More Reliable Pseudo-labels, Better Performance: A Generalized Approach to Single Positive Multi-label Learning
Luong Tran (FPT Software AI Center), Van Nguyen (FPT Software AI Center)
ClassificationVision Language ModelImage
🎯 What it does: A framework for single positive sample multi-label learning, AEVLP, is proposed, which combines a new pseudo-label robust loss (GPR Loss) and dynamic multi-focus pseudo-label generation (DAMP) to achieve high-quality multi-label classification with only a single positive label provided.
Morph: A Motion-free Physics Optimization Framework for Human Motion Generation
Zhuo Li (WeChat, Tencent Inc), Chen Li (WeChat, Tencent Inc)
GenerationOptimizationReinforcement LearningVideoTextPhysics Related
🎯 What it does: The Morph framework is proposed, which combines a motion generator and a physics optimization module to achieve physically feasible motion generation without physical constraints.
MorphoGen: Efficient Unconditional Generation of Long-Range Projection Neuronal Morphology via a Global-to-Local Framework
Tianfang Zhu (Wuhan National Laboratory for Optoelectronics), Anan Li (Hust Suzhou Institute for Brainsmatics)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelPoint Cloud
🎯 What it does: This work proposes an unconditional long-range projecting neuron morphology generation framework called MorphoGen, which adopts a top-down global-to-local generation strategy.
MOSAIC: Generating Consistent, Privacy-Preserving Scenes from Multiple Depth Views in Multi-Room Environments
Zhixuan Liu (Carnegie Mellon University), Jean Oh (Carnegie Mellon University)
GenerationData SynthesisSafty and PrivacyDiffusion modelImage
🎯 What it does: This paper proposes a training-free multi-view diffusion framework called MOSAIC, which utilizes depth maps to generate privacy-friendly RGB images in multi-room indoor environments while ensuring consistency across views.
MosaicDiff: Training-free Structural Pruning for Diffusion Model Acceleration Reflecting Pretraining Dynamics
Bowei Guo (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: A training-free structured pruning method named MosaicDiff is proposed, which can perform pruning with different sparsity levels corresponding to the segmented learning rates of the pre-training of diffusion models, thereby accelerating the sampling process.
MOSCATO: Predicting Multiple Object State Change Through Actions
Parnian Zameni (Northeastern University), Ehsan Elhamifar (Northeastern University)
Object DetectionObject TrackingConvolutional Neural NetworkLarge Language ModelVision Language ModelVideoBenchmark
🎯 What it does: This paper proposes the Multi-Object State Prediction (MOSP) task and the MOSCATO benchmark, and constructs a weakly supervised framework that utilizes pseudo-labeling to learn the evolution of multi-object states across long videos with multiple actions.
MoSiC: Optimal-Transport Motion Trajectory for Dense Self-Supervised Learning
Mohammadreza Salehi (University of Amsterdam), Yuki M Asano (Universidad Tecnológica Nacional)
Object TrackingSegmentationRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: Proposes the MoSiC method, which utilizes long-range point tracking in videos and Sinkhorn transfer clustering to achieve spatiotemporal consistent self-supervised learning of dense features.
Motal: Unsupervised 3D Object Detection by Modality and Task-specific Knowledge Transfer
Hai Wu (Xiamen University), Chenglu Wen (Xiamen University)
Object DetectionAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: Proposes the Motal method to achieve unsupervised 3D object detection, enhancing detection accuracy through multimodal and task-specific knowledge transfer.
Motion Synthesis with Sparse and Flexible Keyjoint Control
Inwoo Hwang (Seoul National University), Young Min Kim (Seoul National University)
GenerationData SynthesisPose EstimationDiffusion modelVideoText
🎯 What it does: A two-stage diffusion-based controllable motion generation framework is proposed, which first generates low-dimensional key joint trajectories using sparse key joint control signals, and then completes natural full-body motion through the complete key joint trajectories, while supporting time-independent control.
Motion-2-to-3: Leveraging 2D Motion Data for 3D Motion Generations
Ruoxi Guo (Zhejiang University), Xiaowei Zhou (Zhejiang University)
GenerationPose EstimationTransformerDiffusion modelVideoText
🎯 What it does: This work proposes the Motion-2-to-3 method, which enhances text-based 3D action generation using 2D action data extracted from 2D videos.
MotionAgent: Fine-grained Controllable Video Generation via Motion Field Agent
Xinyao Liao (Nanyang Technological University), Chi Zhang (Westlake University)
Object DetectionGenerationDepth EstimationTransformerLarge Language ModelDiffusion modelSimultaneous Localization and MappingOptical FlowImageVideoTextBenchmark
🎯 What it does: This paper proposes MotionAgent, which utilizes LLM and object detection to parse motion information from text, generating object trajectories and camera extrinsics, and then achieves text-driven fine-grained image-to-video generation through optical flow synthesis and an optical flow adapter.
MotionCtrl: A Real-time Controllable Vision-Language-Motion Model
Bin Cao (Institute of Automation, Chinese Academy of Sciences), Zongqing Lu (Peking University)
GenerationRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: This study proposes MotionCtrl, a large-scale real-time controllable visual language-action model based on LLaMA-2 7B, and constructs the HuMo100M large-scale multimodal action dataset.
MotionDiff: Training-free Zero-shot Interactive Motion Editing via Flow-assisted Multi-view Diffusion
Yikun Ma, Zhi Jin
GenerationDiffusion modelOptical FlowImagePoint Cloud
🎯 What it does: MotionDiff proposes a training-free zero-shot copy flow-assisted diffusion framework that can achieve complex motion editing such as translation, scaling, rotation, and stretching in multi-view static scenes.
MotionFollower: Editing Video Motion via Score-Guided Diffusion
Shuyuan Tu (Fudan University), Yu-Gang Jiang (Fudan University)
Image TranslationGenerationPose EstimationDiffusion modelScore-based ModelVideo
🎯 What it does: This paper proposes MotionFollower, a video motion editing framework based on diffusion models, which can transfer the motion of a target video to a source video while keeping the background, character appearance, and camera motion of the source video unchanged.
MotionLab: Unified Human Motion Generation and Editing via the Motion-Condition-Motion Paradigm
Ziyan Guo (Singapore University of Technology and Design), Na Zhao (Singapore University of Technology and Design)
GenerationData SynthesisTransformerRectified FlowVideoMultimodality
🎯 What it does: Proposes the MotionLab framework, which unifies the generation and editing of human actions;
MotionShot: Adaptive Motion Transfer across Arbitrary Objects for Text-to-Video Generation
Yanchen Liu (Harbin Institute of Technology), Wenjie Pei (Tongji University)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: Proposes MotionShot, a training-independent text-to-video motion transfer framework.
MotionStreamer: Streaming Motion Generation via Diffusion-based Autoregressive Model in Causal Latent Space
Lixing Xiao (Zhejiang University), Jingbo Wang (Shanghai AI Laboratory)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideoText
🎯 What it does: Proposes the MotionStreamer framework, enabling online, incremental, text-driven human motion generation.
Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos
Yi Chen (University of Hong Kong), Xihui Liu (University of Hong Kong)
GenerationRobotic IntelligenceTransformerLarge Language ModelAuto EncoderVideo
🎯 What it does: This study proposes Moto, which utilizes Latent Motion Tokens obtained through unsupervised learning as a 'language' for autoregressive generative pre-training on video data, and transfers the learned motion priors to real robot manipulation tasks through collaborative fine-tuning.
Move to Understand a 3D Scene: Bridging Visual Grounding and Exploration for Efficient and Versatile Embodied Navigation
Ziyu Zhu (Tsinghua University), Qing Li (State Key Laboratory of General Artificial Intelligence, BIGAI, China)
Robotic IntelligenceTransformerReinforcement LearningContrastive LearningImagePoint Cloud
🎯 What it does: Proposed the MTU3D framework, which integrates visual induction and active exploration to achieve real-time scene understanding and navigation under RGB-D input;
MOVE: Motion-Guided Few-Shot Video Object Segmentation
Kaining Ying (Fudan University), Henghui Ding (Fudan University)
Object DetectionSegmentationConvolutional Neural NetworkTransformerOptical FlowVideoBenchmark
🎯 What it does: This paper proposes the motion-oriented few-shot video object segmentation (MOVE) task and the corresponding dataset, and designs a baseline method DMA based on motion-appearance separation.
MP-HSIR: A Multi-Prompt Framework for Universal Hyperspectral Image Restoration
Zhehui Wu (Wuhan University), Wei He (Wuhan University)
RestorationSuper ResolutionTransformerPrompt EngineeringImage
🎯 What it does: A multi-prompt (spectral, text, visual) framework MP-HSIR is proposed for integrated hyperspectral image denoising, deblurring, super-resolution, inpainting, dehazing, band completion, and other nine tasks.
MPBR: Multimodal Progressive Bidirectional Reasoning for Open-Set Fine-Grained Recognition
Junfu Tan, Yu Liu
RecognitionTransformerAuto EncoderImageMultimodality
🎯 What it does: A multi-modal evolutionary bidirectional reasoning framework (MPBR) is proposed, which achieves open-set fine-grained recognition by extracting hierarchical visual features in a forward manner and aligning with text priors using conditional VAE in a backward manner.
MPG-SAM 2: Adapting SAM 2 with Mask Priors and Global Context for Referring Video Object Segmentation
Fu Rong (Wuhan University), Lefei Zhang (Horizon Robotics)
Object DetectionSegmentationTransformerVision Language ModelVideoMultimodality
🎯 What it does: A new RVOS framework called MPG-SAM 2 based on SAM 2 is proposed, which utilizes a multimodal encoder to generate video-text aligned features and enhances segmentation through mask priors to generate dense prompts and hierarchical global historical aggregation.
MR-FIQA: Face Image Quality Assessment with Multi-Reference Representations from Synthetic Data Generation
Fu-Zhao Ou (City University of Hong Kong), Sam Kwong (Lingnan University)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: This study constructed a controllable generated synthetic facial image dataset called SynFIQA, and utilized multi-domain reference representations (recognition embeddings, spatial, visual-language) to train and annotate a facial image quality assessment (FIQA) model.
MRGen: Segmentation Data Engine For Underrepresented MRI Modalities
Haoning Wu (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
SegmentationGenerationData SynthesisDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a controllable MRI data synthesis engine, MRGen, based on diffusion models, aimed at generating high-quality training data for low-frequency MRI modalities that lack annotations, thereby improving the performance of segmentation models on these modalities.
MS3D: High-Quality 3D Generation via Multi-Scale Representation Modeling
Guan Luo (Tsinghua University), Jianfeng Zhang (ByteDance Seed)
GenerationTransformerAuto EncoderMesh
🎯 What it does: A multi-scale 3D reconstruction framework named MS3D is proposed, capable of generating high-quality textured meshes from sparse viewpoint images or text, achieving progressive refinement through hierarchical sparse structure latent representation and multi-scale image encoders.
MSA2: Multi-task Framework with Structure-aware and Style-adaptive Character Representation for Open-set Chinese Text Recognition
Yangfu Li (East China Normal University), Yue Lu (East China Normal University)
RecognitionContrastive LearningText
🎯 What it does: A multi-task framework MSA-2 is proposed for open-set Chinese text recognition.
MSQ: Memory-Efficient Bit Sparsification Quantization
Seokho Han (Sungkyunkwan University), Jong Hwan Ko (Sungkyunkwan University)
CompressionOptimizationConvolutional Neural NetworkImage
🎯 What it does: A memory-efficient bit-level sparse quantization method MSQ is proposed, which can directly sparsify the least significant bits (LSB) of weights during training, achieving mixed-precision quantization.
MUG: Pseudo Labeling Augmented Audio-Visual Mamba Network for Audio-Visual Video Parsing
Langyu Wang (Institute of Automation, Chinese Academy of Sciences), Jinqiao Wang (Institute of Automation, Chinese Academy of Sciences)
RecognitionSegmentationConvolutional Neural NetworkContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes a pseudo-label enhanced audio-visual Mamba network (MUG), which improves the accuracy of audio and visual video parsing through cross-modal random combination data augmentation and text semantic interaction.