CVPR 2025 Papers — Page 16
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
LP-Diff: Towards Improved Restoration of Real-World Degraded License Plate
Haoyan Gong (Xi'an Jiaotong-Liverpool University), Hongbin Liu (Xi'an Jiaotong-Liverpool University)
RecognitionRestorationDiffusion modelImage
🎯 What it does: This paper proposes the LP-Diff network for license plate image restoration based on diffusion models, as well as the first real multi-frame degraded license plate dataset MDLP, addressing the reconstruction and recognition issues of severely degraded license plates in real-world scenarios.
LPOSS: Label Propagation Over Patches and Pixels for Open-vocabulary Semantic Segmentation
Vladan Stojnić (Czech Technical University in Prague), Giorgos Tolias (NAVER LABS Europe)
SegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: An open-source, training-free vocabulary semantic segmentation method LPOSS and LPOSS+ is proposed, which refines the initial predictions generated by VLMs (such as CLIP) at the patch and pixel levels through label propagation.
LSceneLLM: Enhancing Large 3D Scene Understanding Using Adaptive Visual Preferences
Hongyan Zhi (South China University of Technology), Chuang Gan (UMass Amherst)
TransformerLarge Language ModelVision Language ModelPoint CloudBenchmark
🎯 What it does: The LSceneLLM framework is proposed, which utilizes the visual preferences of LLM to automatically locate task-relevant areas and amplifies and fuses fine-grained features through a pluggable scene amplification module, thereby enhancing the understanding capability of large scene 3D visual language models.
LSNet: See Large, Focus Small
Ao Wang (Tsinghua University), Guiguang Ding (Tsinghua University)
ClassificationObject DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A new lightweight visual network design, LSNet, is proposed, which adopts the 'See Large, Focus Small' strategy, combining large kernel perception and small kernel aggregation to improve the efficiency and accuracy of visual information processing.
LT3SD: Latent Trees for 3D Scene Diffusion
Quan Meng (Technical University of Munich), Angela Dai (Technical University of Munich)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelPoint CloudMesh
🎯 What it does: This paper proposes a hierarchical latent tree structure-based 3D scene diffusion model (LT3SD) that can generate large-scale, infinitely expandable 3D scenes layer by layer from low-frequency geometry to high-frequency details in a patch-by-patch manner.
LUCAS: Layered Universal Codec Avatars
Di Liu (Meta), Chen Cao (Meta)
GenerationCompressionAuto EncoderGaussian SplattingMesh
🎯 What it does: LUCAS is proposed, a general codec avatar model based on a layered structure that separates the face and hair into independent meshes, achieving independent deformation of the face and hair, supporting real-time rendering and high-fidelity Gaussian rendering.
Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment
Ziteng Cui (University of Tokyo), Tatsuya Harada (RIKEN AIP)
RestorationGenerationData SynthesisNeural Radiance FieldGaussian SplattingImage
🎯 What it does: To address the high-quality view synthesis problem in multi-view scenes under low light, overexposure, and varying exposure conditions, a method is proposed that incorporates view-adaptive color matrix mapping and curve adjustment within the 3D Gaussian Splatting framework.
LumiNet: Latent Intrinsics Meets Diffusion Models for Indoor Scene Relighting
Xiaoyan Xing (University of Amsterdam), Anand Bhattad (Toyota Technological Institute at Chicago)
Image TranslationGenerationData SynthesisDiffusion modelFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: A framework for indoor scene re-coloring called LumiNet is proposed, which combines latent intrinsic features with diffusion models, capable of transferring the lighting of one image to another while preserving geometry and material.
Lux Post Facto: Learning Portrait Performance Relighting with Conditional Video Diffusion and a Hybrid Dataset
Yiqun Mei (Netflix Eyeline Studios), Paul Debevec (Johns Hopkins University)
RestorationGenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: By modifying the Stable Video Diffusion model and incorporating an HDR lighting injection mechanism, high-fidelity and temporally stable portrait relighting in videos is achieved.
M-LLM Based Video Frame Selection for Efficient Video Understanding
Kai Hu (Carnegie Mellon University), Trishul Chilimbi (Amazon)
RecognitionRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVideoMultimodality
🎯 What it does: A lightweight multimodal large language model (M-LLM) driven frame selector is proposed, which adaptively selects key frames from videos based on question awareness, thereby enhancing video question-answering performance.
M^3-VOS: Multi-Phase, Multi-Transition, and Multi-Scenery Video Object Segmentation
Zixuan Chen (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Jiao Tong University)
SegmentationVideoBenchmark
🎯 What it does: A multi-stage, multi-transformation, multi-scenario video object segmentation benchmark M-VOS 3 has been established, and a reverse refinement module ReVOS has been proposed on this benchmark to enhance segmentation performance under phase transformations.
M3amba: Memory Mamba is All You Need for Whole Slide Image Classification
Tingting Zheng, Hongxun Yao
ClassificationConvolutional Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: A memory-driven model based on Mamba, called M3amba, is proposed for multi-instance learning classification of whole slide images.
M3GYM: A Large-Scale Multimodal Multi-view Multi-person Pose Dataset for Fitness Activity Understanding in Real-world Settings
Qingzheng Xu (University of Queensland), Xin Yu (University of Queensland)
Pose EstimationTransformerSupervised Fine-TuningVideoMultimodalityMesh
🎯 What it does: A large-scale multimodal, multi-view, multi-person human pose dataset M3GYM has been constructed, recording fitness activities in real gyms and providing fine-grained annotations.
MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction
Xiaohao Xu (University of Michigan), Xiaonan Huang (University of Michigan)
Pose EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a real-time multi-agent collaborative localization and 3D reconstruction framework named MAC-Ego3D, which achieves collaborative pose estimation and high-fidelity dense reconstruction through multi-agent Gaussian consensus.
MAD: Memory-Augmented Detection of 3D Objects
Ben Agro (Waabi), Raquel Urtasun (Waabi)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: A long-term memory module (MAD) is added to the existing 3D object detection model, utilizing the fusion of past predicted trajectories and detection results to improve the detection accuracy of occluded or distant objects.
MaDCoW: Marginal Distortion Correction for Wide-Angle Photography with Arbitrary Objects
Kevin Zhang (University of Maryland), Aaron Hertzmann (Adobe Research)
RestorationOptimizationImage
🎯 What it does: This paper proposes a method called MaDCoW, which corrects distortion of edge objects in wide-angle photography through user-labeled ROIs and lines.
MAGE : Single Image to Material-Aware 3D via the Multi-View G-Buffer Estimation Model
Haoyuan Wang (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
GenerationData SynthesisDiffusion modelImageMesh
🎯 What it does: Predict geometric information and physical material properties (Albedo, Roughness, Metallic) from a single input image to generate a 3D mesh with realistic lighting.
MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM
Vladimir Yugay (University of Amsterdam), Martin R. Oswald (University of Amsterdam)
OptimizationComputational EfficiencyRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a multi-agent SLAM system called MAGiC-SLAM, which supports an arbitrary number of collaborative agents while simultaneously achieving 3D Gaussian scene reconstruction and new view synthesis.
MagicArticulate: Make Your 3D Models Articulation-Ready
Chaoyue Song (Nanyang Technological University), Guosheng Lin (ByteDance Seed)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelMesh
🎯 What it does: Automatically convert static 3D models into animatable models with skeletons and skinning weights.
MagicQuill: An Intelligent Interactive Image Editing System
Zichen Liu (Hong Kong University of Science and Technology), Yujun Shen (Ant Group)
GenerationData SynthesisLarge Language ModelDiffusion modelImageMultimodality
🎯 What it does: MagicQuill has been designed and implemented as an interactive image editing system based on diffusion models, supporting three intuitive brushes: adding, deleting, and coloring. It automatically infers user editing intentions through a multimodal large language model (MLLM), enabling real-time precise editing without the need for manual prompt input.
Magma: A Foundation Model for Multimodal AI Agents
Jianwei Yang (Microsoft Research), Jianfeng Gao (Microsoft Research)
Robotic IntelligenceConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelImageVideoTextMultimodality
🎯 What it does: Train a foundational model MAGE that can understand multimodal inputs and perform actions in both digital and physical environments, possessing perception, understanding, and action capabilities.
Maintaining Consistent Inter-Class Topology in Continual Test-Time Adaptation
Chenggong Ni (Suzhou University of Science and Technology), Tao Zhou (North Minzu University)
ClassificationDomain AdaptationContrastive LearningImageBenchmark
🎯 What it does: A continuous testing adaptive method named Topological Consistency Adaptation (TCA) is proposed to continuously and stably adapt to the changing target domain without accessing the source data.
MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration
Boyun Li (Sichuan University), Xi Peng (Sichuan University)
RestorationSuper ResolutionImage
🎯 What it does: This paper proposes a new Mamba structure called MaIR for image restoration tasks;
Make It Count: Text-to-Image Generation with an Accurate Number of Objects
Lital Binyamin (Bar-Ilan University), Gal Chechik (NVIDIA)
Object DetectionGenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Proposes the CountGen method, which locates and counts object instances in the self-attention features of SDXL, uses the ReLayout network to correct the layout, and then optimizes the generation of images with accurate object counts through layout-guided reasoning.
Make-It-Animatable: An Efficient Framework for Authoring Animation-Ready 3D Characters
Zhiyang Guo (University of Science and Technology of China), Ran Zhang (University of Science and Technology of China)
GenerationPose EstimationComputational EfficiencyTransformerAuto EncoderGaussian SplattingPoint CloudMesh
🎯 What it does: A unified, end-to-end data-driven framework is proposed that can generate high-quality skeletons, weights, and pose-to-rest transitions for 3D humanoid characters of any pose and shape in less than one second, making them immediately animatable.
Making Old Film Great Again: Degradation-aware State Space Model for Old Film Restoration
Yudong Mao (City University of Hong Kong), Shiqi Wang (City University of Hong Kong)
RestorationOptical FlowVideoBenchmark
🎯 What it does: A dynamic degradation-aware restoration framework MambaOFR based on Mamba is proposed to address various mixed degradation issues in old films.
Mamba as a Bridge: Where Vision Foundation Models Meet Vision Language Models for Domain-Generalized Semantic Segmentation
Xin Zhang (National University of Singapore), Robby T. Tan (National University of Singapore)
SegmentationDomain AdaptationTransformerVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes the MFuser framework, which integrates visual foundation models (VFM) and visual language models (VLM) to enhance domain generalization performance in semantic segmentation.
Mamba-Adaptor: State Space Model Adaptor for Visual Recognition
Fei Xie (Shanghai Jiao Tong University), Hongshen Zhao (Southeast University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes Mamba-Adaptor, a pluggable adapter designed to improve the performance of the Mamba state space model in visual tasks.
Mamba-Reg: Vision Mamba Also Needs Registers
Feng Wang (Johns Hopkins University), Cihang Xie (UC Santa Cruz)
ClassificationSegmentationTransformerImage
🎯 What it does: By uniformly inserting registration tokens into the Vision Mamba model and reusing the registration head at the end, high norm abnormal tokens in the feature map are reduced, enhancing classification and segmentation performance.
Mamba4D: Efficient 4D Point Cloud Video Understanding with Disentangled Spatial-Temporal State Space Models
Jiuming Liu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
RecognitionSegmentationVideoPoint Cloud
🎯 What it does: A 4D point cloud video understanding backbone called Mamba4D is proposed based on the State Space Model (Mamba), decoupling space and time, and achieving short-term and long-term spatiotemporal associations through Intra-frame Spatial Mamba and Inter-frame Temporal Mamba.
MambaIC: State Space Models for High-Performance Learned Image Compression
Fanhu Zeng (Tsinghua University), Yan Wang (Tsinghua University)
CompressionImageStochastic Differential Equation
🎯 What it does: MambaIC is proposed, a framework for high-performance learning-based image compression utilizing state space models (SSM), integrating SSM into nonlinear transformations and context models, and incorporating windowed local attention;
MambaIRv2: Attentive State Space Restoration
Hang Guo (Tsinghua University), Yawei Li (ETH Zurich)
RestorationSuper ResolutionCompressionTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes MambaIRv2, an improved state-space model of Mamba, which incorporates ViT-style non-causal modeling and is applied to image super-resolution, denoising, and JPEG compression loss recovery.
MambaOut: Do We Really Need Mamba for Vision?
Weihao Yu (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper questions the necessity of Mamba in visual tasks and proposes a model called MambaOut that removes SSM. It evaluates this model on common visual benchmarks such as ImageNet classification, COCO detection/segmentation, and ADE20K segmentation.
MambaVision: A Hybrid Mamba-Transformer Vision Backbone
Ali Hatamizadeh (NVIDIA), Jan Kautz (NVIDIA)
ClassificationObject DetectionSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a hybrid Mamba-Transformer visual backbone called MambaVision, which redesigns the state space module of Mamba and combines it with self-attention blocks for more efficient visual feature modeling.
MambaVLT: Time-Evolving Multimodal State Space Model for Vision-Language Tracking
Xinqi Liu (Harbin Institute of Technology), Zhenyu He (Pengcheng Laboratory)
Object TrackingTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A visual-language tracking framework called MambaVLT is proposed based on the Mamba state space model, utilizing time-evolving state space to achieve cross-frame feature memory and adaptive updating of reference features.
MambaVO: Deep Visual Odometry Based on Sequential Matching Refinement and Training Smoothing
Shuo Wang (Renmin University of China), Deying Li (Renmin University of China)
Pose EstimationOptimizationSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes a deep visual odometry MambaVO based on the Mamba architecture, achieving real-time high-precision pose estimation through geometric initialization, Mamba sequence matching refinement, and training smoothing techniques.
MammAlps: A Multi-view Video Behavior Monitoring Dataset of Wild Mammals in the Swiss Alps
Valentin Gabeff (Ecole Polytechnique Federale de Lausanne), Devis Tuia (Ecole Polytechnique Federale de Lausanne)
RecognitionObject DetectionSegmentationSupervised Fine-TuningVideoMultimodalityBenchmarkAudio
🎯 What it does: This paper presents the MammAlps dataset, a set of multimodal, multi-view, long-term wildlife behavior monitoring data, which includes video, audio, scene segmentation, and dense hierarchical behavior labels.
MangaNinja: Line Art Colorization with Precise Reference Following
Zhiheng Liu (University of Hong Kong), Ping Luo (Ant Group)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: A reference-guided line art coloring method based on diffusion models, MangaNinja, is proposed, which can automatically align reference images and generate colorful images that are consistent with the line art and rich in details.
Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh
Xiangjun Gao (Hong Kong University of Science and Technology), Long Quan (Hong Kong University of Science and Technology)
Gaussian SplattingMesh
🎯 What it does: This paper proposes a control method based on triangular meshes, achieving free deformation, local editing, and soft body simulation of 3D Gaussian Splatting scenes while maintaining high-quality rendering effects.
ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning
Kailin Li (State Key Laboratory of General Artificial Intelligence), Siyuan Huang (State Key Laboratory of General Artificial Intelligence)
Robotic IntelligenceReinforcement LearningVideo
🎯 What it does: A two-stage MANIPTRANS framework is proposed, first using a large-scale MoCap pre-trained hand trajectory imitation model, and then refining the residual module to achieve precise simulation and task execution of dual robotic hands.
ManiVideo: Generating Hand-Object Manipulation Video with Dexterous and Generalizable Grasping
Youxin Pang (Tsinghua University), Yebin Liu (Tsinghua University)
GenerationData SynthesisRobotic IntelligenceTransformerDiffusion modelVideo
🎯 What it does: Generate spatiotemporally coherent, 3D consistent multi-hand multi-object interaction videos from given motion sequences of hands and objects.
MANTA: A Large-Scale Multi-View and Visual-Text Anomaly Detection Dataset for Tiny Objects
Lei Fan (University of New South Wales), Yang Song (Tsinghua University)
Anomaly DetectionVision Language ModelImageTextMultimodality
🎯 What it does: This paper constructs the MANTA dataset, which includes 137K multi-view images of small objects and corresponding textual knowledge, and provides visual-textual benchmark experiments.
MANTA: Diffusion Mamba for Efficient and Effective Stochastic Long-Term Dense Action Anticipation
Olga Zatsarynna (Motor Europe), Juergen Gall (Lamarr Institute for Machine Learning and Artificial Intelligence)
Diffusion modelVideo
🎯 What it does: A diffusion generative network (MANTA) based on the Mamba state space model is proposed for long-term dense action prediction.
MAP: Unleashing Hybrid Mamba-Transformer Vision Backbone's Potential with Masked Autoregressive Pretraining
Yunze Liu (Tsinghua University), Li Yi (Shanghai Artificial Intelligence Laboratory)
RecognitionObject DetectionSegmentationTransformerContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: A self-supervised pretraining framework named Masked Autoregressive Pretraining (MAP) is proposed for jointly training a hybrid visual backbone network that combines Mamba and Transformer modules.
MAR-3D: Progressive Masked Auto-regressor for High-Resolution 3D Generation
Jinnan Chen (National University of Singapore), Gim Hee Lee (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderMesh
🎯 What it does: A progressive 3D mesh generation framework MAR-3D has been designed and implemented, which combines pyramid VAE and cascading Masked Auto-Regressive (MAR) models to achieve stepwise generation of 3D meshes from low resolution to high resolution.
MARBLE: Material Recomposition and Blending in CLIP-Space
Ta Ying Cheng (University of Oxford), Varun Jampani (Stability AI)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Based on the CLIP space and a pre-trained diffusion model, the MARBLE method is proposed to achieve example-driven transfer of object materials in images, material mixing, and fine-grained attribute control.
MaRI: Material Retrieval Integration across Domains
Jianhui Wang (University of Electronic Science and Technology of China), Jingwei Huang (Tencent Hunyuan3D)
RetrievalTransformerContrastive LearningImage
🎯 What it does: A material retrieval framework named MaRI has been constructed, utilizing a dual DINOv2 encoder to jointly learn a shared embedding space for images and materials, and achieving cross-domain alignment between synthetic and real materials through contrastive learning.
MarkushGrapher: Joint Visual and Textual Recognition of Markush Structures
Lucas Morin (IBM Research), Peter Staar (IBM Research)
RecognitionGenerationData SynthesisTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A multi-modal Markush structure recognition framework called MarkushGrapher is proposed, which can automatically generate graphical representations and variable group tables of Markush structures by utilizing image, text, and layout information simultaneously.
Marten: Visual Question Answering with Mask Generation for Multi-modal Document Understanding
Zining Wang (Meituan), Xiaokang Yang (Shanghai Jiao Tong University)
RecognitionSegmentationOptimizationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a pre-training task based on visual question answering and mask generation called VQAMask, and trains a multimodal large language model named Marten for document-level visual understanding.
MARVEL-40M+: Multi-Level Visual Elaboration for High-Fidelity Text-to-3D Content Creation
Sankalp Sinha (DFKI), Muhammad Zeshan Afzal (BITS Pilani)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelPoint CloudMesh
🎯 What it does: A large-scale 3D annotation dataset called MARVEL-40M+ and a two-stage text-to-3D generation pipeline named MARVEL-FX3D are proposed.
MASH-VLM: Mitigating Action-Scene Hallucination in Video-LLMs through Disentangled Spatial-Temporal Representations
Kyungho Bae (Kyung Hee University), Jinwoo Choi (Kyung Hee University)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: This paper proposes MASH-VLM, a model that alleviates the action-scene illusion in Video-LLM by separating spatial and temporal representations.
Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation
Yongkang Li (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the Mask-Adapter module to improve mask embedding in open vocabulary segmentation, enhancing classification accuracy.
Mask^2DiT: Dual Mask-based Diffusion Transformer for Multi-Scene Long Video Generation
Tianhao Qi (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelVideoText
🎯 What it does: Mask DiT achieves fine-grained text and video segment alignment for long videos in multiple scenes through a dual-mask mechanism, and supports autoregressive scene expansion.
Masked Point-Entity Contrast for Open-Vocabulary 3D Scene Understanding
Yan Wang (State Key Laboratory of General Artificial Intelligence, BIGAI), Siyuan Huang (State Key Laboratory of General Artificial Intelligence, BIGAI)
SegmentationRepresentation LearningContrastive LearningPoint Cloud
🎯 What it does: A framework based on entity contrastive learning, called MPEC, is proposed for open vocabulary 3D scene understanding, which aligns 3D point cloud features with language and achieves unsupervised entity-level representation.
Masked Scene Modeling: Narrowing the Gap Between Supervised and Self-Supervised Learning in 3D Scene Understanding
Pedro Hermosilla (TU Wien), Leon Sick (Ulm University)
SegmentationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: A new 3D scene self-supervised learning framework called Masked Scene Modeling (MSM) is proposed, along with an evaluation protocol designed for hierarchical UNet.
MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks
Yifei Liu (Shanghai AI Laboratory), Xiao Sun (Shanghai AI Laboratory)
CompressionOptimizationGaussian SplattingPoint Cloud
🎯 What it does: The MaskGaussian method is proposed, which treats 3D Gaussian points as probabilistic entities and implements adaptive pruning through masked-rasterization, significantly reducing the number of Gaussians while maintaining image quality.
MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction
Jingcheng Ni (SenseTime Research), Zehuan Wu (SenseTime Research)
GenerationAutonomous DrivingTransformerDiffusion modelRectified FlowAuto EncoderWorld ModelVideo
🎯 What it does: This paper proposes MaskGWM, a driving world model based on DiT, which enhances long-term prediction and multi-view generation performance by incorporating a video mask reconstruction task during training.
Masking meets Supervision: A Strong Learning Alliance
Byeongho Heo (NAVER AI Lab), Dongyoon Han (NAVER AI Lab)
ClassificationSegmentationKnowledge DistillationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: The MaskSub framework is proposed, which runs the main branch and the sub-branch in parallel. The sub-branch uses a high ratio of masked inputs and is guided by self-distillation-style relaxed loss during training, allowing supervised learning to withstand strong masked augmentation.
MaSS13K: A Matting-level Semantic Segmentation Benchmark
Chenxi Xie (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
SegmentationTransformerImageBenchmark
🎯 What it does: A high-resolution semantic segmentation dataset, MaSS13K, has been proposed, and based on this, an efficient Transformer semantic segmentation model, MaSSFormer, designed for 4K images has been developed.
MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors
Riku Murai (Imperial College London), Andrew J. Davison (Imperial College London)
Object TrackingPose EstimationDepth EstimationRobotic IntelligenceSimultaneous Localization and MappingImageVideo
🎯 What it does: A real-time dense monocular SLAM system based on the two-view 3D reconstruction prior MASt3R is proposed;
MatAnyone: Stable Video Matting with Consistent Memory Propagation
Peiqing Yang (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
Object DetectionSegmentationTransformerVideo
🎯 What it does: A target allocation video matting framework called MatAnyone is proposed, which can finely capture edge details while maintaining semantic stability, achieving high-quality and temporally consistent alpha matting.
MAtCha Gaussians: Atlas of Charts for High-Quality Geometry and Photorealism From Sparse Views
Antoine Guedon (University Gustave Eiffel), Ko Nishino (Kyoto University)
GenerationDepth EstimationNeural Radiance FieldSimultaneous Localization and MappingImageMesh
🎯 What it does: The method quickly reconstructs high-quality 3D meshes from sparse RGB images, balancing lighting realism and geometric detail.
MATCHA: Towards Matching Anything
Fei Xue (University of Cambridge), Qunjie Zhou (NVIDIA)
TransformerDiffusion modelContrastive LearningImageVideo
🎯 What it does: MATCHA is proposed, a unified feature model capable of achieving 'matching anything' across geometric, semantic, and temporal matching tasks.
Material Anything: Generating Materials for Any 3D Object via Diffusion
Xin Huang (Northwestern Polytechnical University), Qing Wang (Northwestern Polytechnical University)
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: Automatically generate physically based rendering (PBR) materials for arbitrary 3D meshes, providing end-to-end, ready-to-use material textures.
Matrix-Free Shared Intrinsics Bundle Adjustment
Daniel Safari (Sony Semiconductor Solutions)
OptimizationComputational EfficiencySimultaneous Localization and MappingImage
🎯 What it does: A matrix-free parallel solver for Shared Intrinsic Parameter Bundle Adjustment (SI-BA) is proposed, utilizing adaptive sparse structures, SIMD/CUDA acceleration, and single-precision numerical stability techniques, achieving approximately 8 times the speedup and 10 times the memory compression compared to Ceres.
Matrix3D: Large Photogrammetry Model All-in-One
Yuanxun Lu (Nanjing University), Shiwei Li (Apple)
GenerationPose EstimationDepth EstimationTransformerDiffusion modelGaussian SplattingImageMultimodality
🎯 What it does: A unified optical 3D reconstruction model, Matrix3D, has been constructed, capable of simultaneously performing three major photogrammetry tasks: camera pose estimation, depth prediction, and novel view synthesis.
MBQ: Modality-Balanced Quantization for Large Vision-Language Models
Shiyao Li (Tsinghua University), Yu Wang (Tsinghua University)
CompressionOptimizationTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes a post-training quantization method for large visual-language models called MBQ (Modality-Balanced Quantization), which improves quantization quality by considering the different sensitivities to errors of the visual and language modalities during the calibration process.
MC^2: Multi-concept Guidance for Customized Multi-concept Generation
Jiaxiu Jiang (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
GenerationData SynthesisDiffusion modelImageBenchmark
🎯 What it does: Through optimization during inference, MC 2 seamlessly integrates multiple trained single-concept custom models (such as Textual Inversion, LoRA, DreamBooth) to generate multi-concept images.
MCCD: Multi-Agent Collaboration-based Compositional Diffusion for Complex Text-to-Image Generation
Mingcheng Li (Fudan University), Lihua Zhang (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: This paper studies a training-free text-to-image generation framework called MCCD, which utilizes multi-agent collaboration to parse complex prompts and generates high-quality scene images through hierarchical synthesis diffusion.
MDP: Multidimensional Vision Model Pruning with Latency Constraint
Xinglong Sun (NVIDIA), Jose M. Alvarez (NVIDIA)
ClassificationObject DetectionOptimizationTransformerImage
🎯 What it does: A unified multidimensional structure pruning framework MDP is proposed, which can simultaneously prune at the channel, head, embedding dimension, and whole block level, while seeking a globally optimal subnetwork under a given latency budget.
MEAT: Multiview Diffusion Model for Human Generation on Megapixels with Mesh Attention
Yuhan Wang (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
GenerationData SynthesisDiffusion modelImageVideoMesh
🎯 What it does: A multi-view diffusion model MEAT is proposed, capable of generating highly consistent and detail-rich multi-view human images at a resolution of 1024×1024 pixels, with only a front image as input.
MedUnifier: Unifying Vision-and-Language Pre-training on Medical Data with Vision Generation Task using Discrete Visual Representations
Ziyang Zhang (Northwestern University), Si Yong Yeo (Nanyang Technological University)
GenerationTransformerContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: Developed MedUnifier, a unified medical vision-language pre-training framework that integrates text-guided image generation and multimodal learning tasks.
Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering
Liang Chen (Beijing University of Posts and Telecommunications), Yuankai Qi (Macquarie University)
Graph Neural NetworkContrastive LearningImage
🎯 What it does: A multi-scale multi-view clustering framework named Medusa is proposed, achieving superior clustering performance through dual graph diffusion, hypergraph contrastive learning, and fine-grained sample-level fusion.
MEET: Towards Memory-Efficient Temporal Sparse Deep Neural Networks
Zeqi Zhu (Snap Inc), Orlando Moreira (Snap Inc)
Object DetectionSegmentationPose EstimationCompressionComputational EfficiencyNeural Architecture SearchConvolutional Neural NetworkVideo
🎯 What it does: The MEET framework is designed to achieve memory-efficient inference on time-sparse ∆Σ-convolution DNNs. This framework significantly reduces neuron state memory by transforming activation compression into a weight quantization/compression problem and combining it with mixed space-time suppression, enabling the model to be fully deployed on embedded event-driven platforms such as GrAI-VIP.
MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing
Cong Wang (Tsinghua University), Song-Hai Zhang (Tsinghua University)
GenerationData SynthesisGaussian SplattingVideoMesh
🎯 What it does: A hybrid mesh-Gaussian representation for full-body animated avatars, MeGA, is proposed, capable of high-quality rendering of faces and hair while supporting editing.
MEGA: Masked Generative Autoencoder for Human Mesh Recovery
Guénolé Fiche (Naver Labs Europe), Francesc Moreno-Noguer (Amazon)
GenerationPose EstimationTransformerAuto EncoderMesh
🎯 What it does: A method for human mesh recovery from a single RGB image based on Masked Generative Autoencoder (MEGA) is proposed, which can generate diverse human meshes in both deterministic and stochastic modes.
MegaSaM: Accurate, Fast and Robust Structure and Motion from Casual Dynamic Videos
Zhengqi Li (Google), Noah Snavely (Google)
Pose EstimationDepth EstimationOptimizationSimultaneous Localization and MappingOptical FlowVideo
🎯 What it does: We propose MegaSaM, a real-time camera pose and depth estimation pipeline for arbitrary dynamic monocular videos.
MegaSynth: Scaling Up 3D Scene Reconstruction with Synthesized Data
Hanwen Jiang (University of Texas at Austin), Hao Tan (Adobe Research)
RestorationGenerationData SynthesisTransformerPoint CloudMesh
🎯 What it does: This paper presents MegaSynth, a dataset of 700,000 non-semantic 3D scenes, which is used alongside real data to train large reconstruction models (LRM), significantly improving scene-level 3D reconstruction results.
Memories of Forgotten Concepts
Matan Rusanovsky (Tel Aviv University), Shai Avidan (Tel Aviv University)
RestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper studies whether diffusion models truly forget erased concepts after concept erasure, proposing an analytical method to recover erased concept images by finding high-likelihood seeds through inversion.
MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images
Aniruddha Ganguly (Stony Brook University), Chao Chen (Stony Brook University)
Graph Neural NetworkImageBiomedical Data
🎯 What it does: Using multi-level graphs and graph neural networks, combined with spatial and feature clustering, to jointly predict gene expression from whole slide images.
MergeVQ: A Unified Framework for Visual Generation and Representation with Disentangled Token Merging and Quantization
Siyuan Li (Zhejiang University), Zhen Lei (University of Chinese Academy of Sciences)
GenerationRepresentation LearningConvolutional Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes MergeVQ, a unified framework that combines token merging and vector quantization to achieve unified training and inference for visual representation learning and image generation.
MESC-3D:Mining Effective Semantic Cues for 3D Reconstruction from a Single Image
Shaoming Li (Ocean University of China), Zhi Liu (Shandong University)
GenerationData SynthesisTransformerPrompt EngineeringContrastive LearningImagePoint Cloud
🎯 What it does: To address the 3D point cloud reconstruction problem from a single-view image, the authors propose a new method called MESC-3D, which guides point cloud generation through effective semantic cues.
Mesh Mamba: A Unified State Space Model for Saliency Prediction in Non-Textured and Textured Meshes
Kaiwei Zhang (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
Graph Neural NetworkPoint CloudMesh
🎯 What it does: This study proposes the Mesh Mamba model, which uses a unified state space model to predict visual attention (saliency) maps for both textured and non-textured meshes.
MeshArt: Generating Articulated Meshes with Structure-Guided Transformers
Daoyi Gao (Technical University of Munich), Angela Dai (Technical University of Munich)
GenerationData SynthesisTransformerAuto EncoderMesh
🎯 What it does: We propose MeshArt, a hierarchical method based on Transformer for generating 3D mesh models with movable joints;
MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation
Zilong Chen (Tsinghua University), Huaping Liu (Tsinghua University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageMesh
🎯 What it does: This paper presents MeshGen, which can generate high-quality, PBR-textured 3D meshes from a single image.
MET3R: Measuring Multi-View Consistency in Generated Images
Mohammad Asim (Max Planck Institute for Informatics), Jan Eric Lenssen (Max Planck Institute for Informatics)
GenerationData SynthesisImageVideo
🎯 What it does: This paper proposes a new metric called MEt3R to measure the 3D consistency of generated image sequences from different viewpoints.
Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning
Zichen Tian (Singapore Management University), Qianru Sun (Singapore Management University)
OptimizationMeta LearningImage
🎯 What it does: This paper studies and proposes an adaptive parameter-efficient fine-tuning method named MetaPEFT, designed for long-tail tasks in remote sensing and natural images, which can automatically learn the insertion positions, layer depths, and scaling factors of PEFT modules.
METASCENES: Towards Automated Replica Creation for Real-world 3D Scans
Huangyue Yu (State Key Laboratory of General Artificial Intelligence, BIGAI), Siyuan Huang (State Key Laboratory of General Artificial Intelligence, BIGAI)
Data SynthesisRetrievalOptimizationTransformerLarge Language ModelVision Language ModelMultimodalityPoint CloudBenchmark
🎯 What it does: A large-scale simulative 3D scene dataset called METASCENES is proposed, along with the development of the SCAN2SIM multimodal alignment model, which automates the replacement of objects in real scanned scenes with high-quality interactive 3D assets. Additionally, two evaluation benchmarks for micro-scene synthesis and cross-domain VLN are provided.
MetaShadow: Object-Centered Shadow Detection, Removal, and Synthesis
Tianyu Wang (Adobe Research), Soo Ye Kim (Adobe Research)
Image TranslationRestorationGenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImageVideo
🎯 What it does: Proposes the MetaShadow framework, which unifies the implementation of object-centered shadow detection, removal, and synthesis.
MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning
Wenhao Gu (Concordia University), Yang Wang (Concordia University)
RecognitionMeta LearningTransformerPrompt EngineeringAuto EncoderImage
🎯 What it does: A visual prompt tuning framework based on meta-learning is proposed, which quickly personalizes handwritten text recognition models with a small number of unlabeled handwritten samples during testing.
MetricGrids: Arbitrary Nonlinear Approximation with Elementary Metric Grids based Implicit Neural Representation
Shu Wang (Shandong University), Jinglin Zhang (Shandong University)
RestorationGenerationSuper ResolutionNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes MetricGrids, which utilizes multi-dimensional metric grids to approximate continuous signals through Taylor expansion, achieving high-order nonlinear feature interpolation.
MExD: An Expert-Infused Diffusion Model for Whole-Slide Image Classification
Jianwei Zhao (University of Electronic Science and Technology of China), Huazhu Fu
ClassificationTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: An expert-integrated diffusion model for whole slide image classification (MExD) is proposed, which filters key information through a dynamic Mixture-of-Experts (Dyn-MoE) aggregator and directly generates class distributions using a diffusion classifier.
MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting
Mengqiu Xu (Beijing University of Posts and Telecommunications), Jun Guo (Beijing University of Posts and Telecommunications)
Convolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: A global multi-region multi-satellite sea fog detection and prediction dataset, MFogHub, has been constructed, and multi-model benchmark experiments have been conducted on it.
MG-MotionLLM: A Unified Framework for Motion Comprehension and Generation across Multiple Granularities
Bizhu Wu (Shenzhen University), Linlin Shen (Shenzhen University)
GenerationData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: A unified multi-granularity motion-language model MG-MotionLLM is proposed, capable of performing motion understanding and generation from coarse text to fine-grained motion scripts within the same framework, supporting various tasks such as Text-to-Motion, Motion-to-Text, and Motion-to-Detailed-Text.
MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism
Zhixiong Nan (Chongqing University), Tao Xiang (Chongqing University)
Object DetectionTransformerImage
🎯 What it does: A parallel multi-query (MI) decoder architecture is proposed to enhance the feature utilization of DETR-based object detection models.
MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing
Feifei Shao (Zhejiang University), Jun Xiao (Zhejiang University)
RestorationSegmentationTransformerPrompt EngineeringPoint Cloud
🎯 What it does: A multi-granularity context adaptive sampling framework MICAS is proposed for multi-task In-Context Learning of 3D point clouds, addressing the sampling sensitivity issues both between and within tasks.
MicroVQA: A Multimodal Reasoning Benchmark for Microscopy-Based Scientific Research
James Burgess (Stanford University), Serena Yeung-Levy (Chan Zuckerberg Biohub Network)
Large Language ModelAgentic AIPrompt EngineeringImageMultimodalityBenchmark
🎯 What it does: This paper presents MicroVQA, a multimodal visual question answering benchmark based on biological microscope images, aimed at evaluating expert-level image understanding, hypothesis generation, and experimental design reasoning capabilities.
MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation
Zehuan Huang, Lu Sheng
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImagePoint Cloud
🎯 What it does: MIDI is proposed, a multi-instance diffusion model that can generate multi-instance 3D scenes from a single image while maintaining accurate spatial relationships.
Mimic In-Context Learning for Multimodal Tasks
Yuchu Jiang (Southeast University), Xu Yang (Southeast University)
RecognitionGenerationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the MimIC method, which simulates the effects of In-Context Learning using lightweight learnable offset vectors in each attention head of a multimodal Transformer, without the need to carry examples during inference.
Mimir: Improving Video Diffusion Models for Precise Text Understanding
Shuai Tan (Ant Group), Ming Yang (Ant Group)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoText
🎯 What it does: Mimir achieves precise text understanding in video diffusion models by introducing large language models and text encoders, and generates high-quality videos through Token Fuser for feature fusion.