CVPR 2023 Papers — Page 14
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
Micron-BERT: BERT-Based Facial Micro-Expression Recognition
Xuan-Bac Nguyen (University of Arkansas), Khoa Luu (University of Arkansas)
RecognitionTransformerSupervised Fine-TuningVideo
🎯 What it does: This paper proposes Micron-BERT, a micro-expression recognition framework based on self-supervised BERT, capable of automatically capturing and locating subtle facial micro-movements.
MIME: Human-Aware 3D Scene Generation
Hongwei Yi (Max Planck Institute for Intelligent Systems), Michael J. Black (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisTransformerPoint CloudMesh
🎯 What it does: Generate an indoor furniture layout compatible with 3D human motion and a blank plane, forming a complete 3D scene.
Mind the Label Shift of Augmentation-Based Graph OOD Generalization
Junchi Yu (Chinese Academy of Sciences), Ran He (Chinese Academy of Sciences)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the label shift problem in out-of-distribution generalization using graph neural networks (GNNs) and proposes a label-invariant subgraph augmentation method called LiSA based on a variable subgraph generator, which generates diverse environments to learn invariant GNNs.
Minimizing Maximum Model Discrepancy for Transferable Black-Box Targeted Attacks
Anqi Zhao (University of Electronic Science and Technology of China), Lixin Duan (University of Electronic Science and Technology of China)
Adversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This study investigates black-box targeted attacks and proposes theories and algorithms from the perspective of model differences.
Minimizing the Accumulated Trajectory Error To Improve Dataset Distillation
Jiawei Du (Agency for Science Technology and Research), Haizhou Li (Chinese University of Hong Kong)
Data SynthesisOptimizationKnowledge DistillationImage
🎯 What it does: This paper proposes a Flat Trajectory Distillation (FTD) method aimed at reducing the cumulative trajectory error during the dataset distillation process, thereby improving the performance of the synthetic dataset in actual training.
MISC210K: A Large-Scale Dataset for Multi-Instance Semantic Correspondence
Yixuan Sun (Fudan University), Wenqiang Zhang (Fudan University)
Object DetectionSegmentationTransformerImage
🎯 What it does: This paper proposes a multi-instance semantic correspondence task and constructs the MISC210K dataset, which consists of 218K image pairs and 34 categories. A dual-path collaborative learning framework (instance-level co-segmentation and keypoint matching) is designed to address this task.
MIST: Multi-Modal Iterative Spatial-Temporal Transformer for Long-Form Video Question Answering
Difei Gao (National University of Singapore), Mike Zheng Shou (National University of Singapore)
RecognitionRetrievalComputational EfficiencyTransformerVision Language ModelVideoMultimodality
🎯 What it does: The MIST model is proposed, which achieves efficient reasoning of multiple events and multi-scale visual concepts for long-duration video question answering through iterative spatio-temporal attention and hierarchical segment/region selection.
Mitigating Task Interference in Multi-Task Learning via Explicit Task Routing With Non-Learnable Primitives
Chuntao Ding (Beijing Jiaotong University), Vishnu Naresh Boddeti
ClassificationSegmentationDepth EstimationAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: A multi-task network ETR-NLP based on non-learnable primitives and explicit task routing is proposed to alleviate task interference.
Mixed Autoencoder for Self-Supervised Visual Representation Learning
Kai Chen (Hong Kong University of Science and Technology), Dit-Yan Yeung (Hong Kong University of Science and Technology)
Object DetectionSegmentationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: A Mixed Autoencoder (MixedAE) is proposed, which introduces image mixing and homogenous recognition tasks based on MAE to enhance self-supervised visual representation.
MixMAE: Mixed and Masked Autoencoder for Efficient Pretraining of Hierarchical Vision Transformers
Jihao Liu (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Object DetectionSegmentationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: This paper proposes a Mixed and Masked Autoencoder (MixMAE), which enhances the efficiency of self-supervised visual representation learning by randomly mixing visible patches from two images and training without the [MASK] symbol in the encoder, followed by dual reconstruction of the two original images in the decoder.
MixNeRF: Modeling a Ray With Mixture Density for Novel View Synthesis From Sparse Inputs
Seunghyeon Seo (Seoul National University), Nojun Kwak (Seoul National University)
GenerationData SynthesisDepth EstimationNeural Radiance FieldImageBenchmark
🎯 What it does: This paper proposes MixNeRF, a method for training neural radiance fields (NeRF) under sparse input conditions. It models the color and density along rays using a mixture density model and incorporates ray depth estimation and weight regeneration as auxiliary training.
MixPHM: Redundancy-Aware Parameter-Efficient Tuning for Low-Resource Visual Question Answering
Jingjing Jiang (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
TransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: This paper proposes MixPHM, a redundancy-aware parameter-efficient fine-tuning method for optimizing large-scale vision-language models (VLM) in low-resource visual question answering (VQA) tasks, achieving better results than full fine-tuning.
MixSim: A Hierarchical Framework for Mixed Reality Traffic Simulation
Simon Suo (Waabi), Raquel Urtasun (Waabi)
Autonomous DrivingRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningGraphTime Series
🎯 What it does: Designed the MIXSIM framework to achieve controllable and reproducible mixed reality simulations of real-world traffic scenarios;
MixTeacher: Mining Promising Labels With Mixed Scale Teacher for Semi-Supervised Object Detection
Liang Liu (Youtu Lab, Tencent), Chengjie Wang (Tencent)
Object DetectionKnowledge DistillationImage
🎯 What it does: A mixed-scale teacher framework called MixTeacher is proposed in semi-supervised object detection, which utilizes a mixed-scale feature pyramid to enhance the quality of pseudo-labels and explores low-confidence labels through a cross-scale score enhancement mechanism.
MM-3DScene: 3D Scene Understanding by Customizing Masked Modeling With Informative-Preserved Reconstruction and Self-Distilled Consistency
Mingye Xu (Shenzhen Institute of Advanced Technology), Yu Qiao (Shenzhen Institute of Advanced Technology)
Object DetectionSegmentationTransformerAuto EncoderContrastive LearningPoint Cloud
🎯 What it does: The MM-3DScene framework is proposed to achieve self-supervised pre-training for large-scale 3D scenes, utilizing information-preserving masking and self-distillation consistency learning.
MM-Diffusion: Learning Multi-Modal Diffusion Models for Joint Audio and Video Generation
Ludan Ruan (Renmin University of China), Baining Guo (Microsoft Research)
GenerationData SynthesisDiffusion modelAuto EncoderVideoMultimodalityAudio
🎯 What it does: The MM-Diffusion model is proposed to achieve joint generation of audio and video under unconditional and zero-shot conditions.
MMANet: Margin-Aware Distillation and Modality-Aware Regularization for Incomplete Multimodal Learning
Shicai Wei (University of Electronic Science and Technology of China), Yang Luo (University of Electronic Science and Technology of China)
RecognitionSegmentationKnowledge DistillationImageMultimodality
🎯 What it does: The MMANet framework is proposed, which jointly achieves robust learning for missing modalities through a teacher network, a deployment network, and a regularization network, and enhances model performance through Marginal-Aware Distillation (MAD) and Modality-Aware Regularization (MAR).
MMG-Ego4D: Multimodal Generalization in Egocentric Action Recognition
Xinyu Gong (Meta Reality Labs), Rakesh Ranjan (University of Texas at Austin)
RecognitionTransformerVideoMultimodalityBenchmarkAudio
🎯 What it does: This study proposes the problem of multimodal generalization (MMG) for first-person action recognition and constructs the MMG-Ego4D dataset and benchmark, which includes video, audio, and IMU three modalities.
MMVC: Learned Multi-Mode Video Compression With Block-Based Prediction Mode Selection and Density-Adaptive Entropy Coding
Bowen Liu (University of Michigan), Hun-Seok Kim (University of Michigan)
CompressionConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: A multi-mode video compression framework MMVC based on block-level adaptive mode selection is proposed, combining ConvLSTM, optical flow conditional feature prediction, feature propagation, and skip modes, along with channel pruning and density-adaptive entropy coding.
Mobile User Interface Element Detection via Adaptively Prompt Tuning
Zhangxuan Gu (Ant Group), Weiqiang Wang (Ant Group)
RecognitionObject DetectionConvolutional Neural NetworkPrompt EngineeringContrastive LearningImageMultimodality
🎯 What it does: Proposes the Adaptively Prompt Tuning (APT) module, which dynamically adjusts category prompts based on OCR descriptions and visual features, thereby enhancing the performance of CLIP-based MUI element detection.
MobileBrick: Building LEGO for 3D Reconstruction on Mobile Devices
Kejie Li (University of Oxford), Victor Adrian Prisacariu (University of Oxford)
Data SynthesisDepth EstimationNeural Radiance FieldImagePoint CloudBenchmark
🎯 What it does: Created the MobileBrick dataset and conducted benchmark evaluations on multi-view reconstruction, view synthesis, and depth map enhancement.
MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures
Zhiqin Chen (Google Research), Andrea Tagliasacchi (Simon Fraser University)
GenerationComputational EfficiencyNeural Radiance FieldMesh
🎯 What it does: Achieve efficient real-time rendering of neural radiance fields on mobile devices by representing NeRF as textured polygon meshes and using a small MLP in the traditional z-buffer rasterization pipeline.
MobileOne: An Improved One Millisecond Mobile Backbone
Pavan Kumar Anasosalu Vasu (Apple), Anurag Ranjan (Apple)
Object DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Proposes the MobileOne architecture, achieving inference speeds below 1 ms on mobile devices while maintaining high accuracy.
MobileVOS: Real-Time Video Object Segmentation Contrastive Learning Meets Knowledge Distillation
Roy Miles (Samsung Research UK), Albert Saà-Garriga (Samsung Research UK)
SegmentationKnowledge DistillationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: We propose MobileVOS, a framework for real-time semi-supervised video object segmentation on mobile devices, which utilizes knowledge distillation and supervised contrastive learning to transfer knowledge from a large teacher network to a small student network.
Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners
Zitian Chen (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
TransformerMixture of ExpertsImage
🎯 What it does: Introducing Mod-Squad—a modular multi-task learning framework based on Vision Transformer, which embeds Mixture-of-Experts (MoE) in the Transformer layers and facilitates collaboration and specialization between tasks through task embeddings.
Modality-Agnostic Debiasing for Single Domain Generalization
Sanqing Qu (Tongji University), Tao Mei (HiDream.ai Inc.)
SegmentationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImageTextPoint Cloud
🎯 What it does: A modality-agnostic debiasing framework MAD is designed, using two classifiers (bias branch and general branch) to enhance single-source domain generalization performance.
Modality-Invariant Visual Odometry for Embodied Vision
Marius Memmel (University of Washington), Amir Zamir (Swiss Federal Institute of Technology)
TransformerSimultaneous Localization and MappingImageMultimodality
🎯 What it does: A Transformer-based visual odometry model VOT is proposed, which can maintain positioning performance even when various optional sensors (RGB, Depth, etc.) are missing;
MoDAR: Using Motion Forecasting for 3D Object Detection in Point Cloud Sequences
Yingwei Li (Waymo LLC), Dragomir Anguelov (Waymo LLC)
Object DetectionAutonomous DrivingTransformerSimultaneous Localization and MappingPoint CloudSequential
🎯 What it does: This paper achieves 3D object detection based on sequential data by fusing virtual points generated from motion prediction (MoDAR) with LiDAR point clouds.
Model Barrier: A Compact Un-Transferable Isolation Domain for Model Intellectual Property Protection
Lianyu Wang (Institute of High Performance Computing), Huazhu Fu (Nanjing University of Aeronautics and Astronautics)
Domain AdaptationSafty and PrivacyConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the Compact Un-Transferable Isolation Domain (CUTI-domain), which enhances private style features within the authorized domain and constructs similar isolation domains to limit the model's transfer and recognition capabilities in unauthorized domains, thereby achieving model IP protection.
Model-Agnostic Gender Debiased Image Captioning
Yusuke Hirota (Osaka University), Noa Garcia (Osaka University)
GenerationTransformerLarge Language ModelImageText
🎯 What it does: This paper proposes a general framework named LIBRA, aimed at simultaneously reducing two types of gender bias (context→gender and gender→context) in image captioning models by synthesizing biased captions and training a debiasing caption generator.
Modeling Entities As Semantic Points for Visual Information Extraction in the Wild
Zhibo Yang (Huazhong University of Science and Technology), Cong Yao (Alibaba Group)
RecognitionObject DetectionConvolutional Neural NetworkContrastive LearningImageTextMultimodality
🎯 What it does: Proposed the ESP framework to achieve end-to-end visual information extraction, unifying entity localization, annotation, and linking, and released the challenging SIBR dataset;
Modeling Inter-Class and Intra-Class Constraints in Novel Class Discovery
Wenbin Li (Nanjing University), Yang Gao (Nanjing University)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a single-stage learning framework designed specifically for the task of Novel Class Discovery, based on symmetric KL divergence with cross-class and internal consistency constraints;
Modeling the Distributional Uncertainty for Salient Object Detection Models
Xinyu Tian (Northwestern Polytechnical University), Yuchao Dai (Northwestern Polytechnical University)
Object DetectionImage
🎯 What it does: This paper studies the distribution uncertainty of salient object detection models and proposes a method to quantify OOD samples using long-tail learning, single model uncertainty estimation, and testing strategies.
Modeling Video As Stochastic Processes for Fine-Grained Video Representation Learning
Heng Zhang (Renmin University of China), Bing Su (Renmin University of China)
RetrievalRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: Modeling videos as stochastic processes, utilizing Brownian Bridge to constrain the temporal evolution of frame-level representations;
Modernizing Old Photos Using Multiple References via Photorealistic Style Transfer
Agus Gunawan (KAIST), Munchurl Kim (KAIST)
Image TranslationRestorationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A multi-reference image modernization framework MROPM is proposed, which modernizes old photos through lighting style transfer and enhancement in one go.
MoDi: Unconditional Motion Synthesis From Diverse Data
Sigal Raab (Tel Aviv University), Daniel Cohen-Or (Tel Aviv University)
GenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkVideo
🎯 What it does: An unsupervised motion generation model MoDi has been constructed, capable of learning motion priors from extremely diverse, unstructured, and unlabeled datasets, and achieving inverse mapping of real motion through an encoder, supporting various motion editing and generation tasks.
Modular Memorability: Tiered Representations for Video Memorability Prediction
Théo Dumont (Mines Paris PSL Research University), Camilo L. Fosco (Memorable AI)
RecognitionSegmentationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: A modular short-term video memorability prediction model M3-S has been constructed, utilizing four modules: low-level visual features, scene segmentation, action recognition, and contextual similarity to obtain hierarchical memory features, which are then fused for regression prediction.
Mofusion: A Framework for Denoising-Diffusion-Based Motion Synthesis
Rishabh Dabral (Max Planck Institute for Informatics), Christian Theobalt (Max Planck Institute for Informatics)
GenerationData SynthesisDiffusion modelTextMultimodalityAudio
🎯 What it does: Proposes the MoFusion framework, which utilizes denoising diffusion models to achieve 3D human motion synthesis based on text or music, and supports interactive editing.
MoLo: Motion-Augmented Long-Short Contrastive Learning for Few-Shot Action Recognition
Xiang Wang (Huazhong University of Science and Technology), Nong Sang (Huazhong University of Science and Technology)
RecognitionTransformerContrastive LearningVideo
🎯 What it does: A MoLo framework for few-shot action recognition is proposed, which integrates global temporal information and motion details for fine matching.
MonoATT: Online Monocular 3D Object Detection With Adaptive Token Transformer
Yunsong Zhou (Shanghai Jiao Tong University), Minyi Guo (Shanghai Jiao Tong University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkTransformerPoint CloudBenchmark
🎯 What it does: An online monocular 3D object detection framework called MonoATT is proposed, which utilizes an adaptive visual Transformer to generate tokens of irregular sizes and shapes to improve detection accuracy and reduce latency.
MonoHuman: Animatable Human Neural Field From Monocular Video
Zhengming Yu (SenseTime Research), Kwan-Yee Lin (Shanghai AI Laboratory)
GenerationPose EstimationNeural Radiance FieldVideo
🎯 What it does: Learning animatable human neural fields from monocular videos to achieve free-viewpoint and arbitrary pose rendering.
MOSO: Decomposing MOtion, Scene and Object for Video Prediction
Mingzhen Sun (Institute of Automation, Chinese Academy of Sciences), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
GenerationData SynthesisTransformerAuto EncoderVideo
🎯 What it does: This paper proposes a two-stage MOSO framework, which first uses MOSO-VQVAE to decompose videos into three types of discrete tokens: motion, scene, and object, and then uses MOSO-Transformer to predict future videos at the token level.
MoStGAN-V: Video Generation With Temporal Motion Styles
Xiaoqian Shen (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkVideo
🎯 What it does: The MoStGAN-V model is proposed, which introduces time-dependent motion styles and the MoStAtt attention mechanism based on StyleGAN-V, achieving more natural and diverse dynamic video generation.
MOT: Masked Optimal Transport for Partial Domain Adaptation
You-Wei Luo, Chuan-Xian Ren
SegmentationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an improved U-Net semantic segmentation network that integrates attention mechanisms and multi-scale features;
Motion Information Propagation for Neural Video Compression
Linfeng Qi (University of Science and Technology of China), Yan Lu (Microsoft Research Asia)
CompressionTransformerOptical FlowVideo
🎯 What it does: A neural video compression framework (DCVC-MIP) has been designed to achieve bidirectional information interaction between motion coding and frame coding, and hybrid context generation has been proposed to better utilize multi-scale motion information.
MotionDiffuser: Controllable Multi-Agent Motion Prediction Using Diffusion
Chiyu “Max” Jiang, Dragomir Anguelov
Autonomous DrivingOptimizationTransformerDiffusion modelPoint CloudOrdinary Differential Equation
🎯 What it does: Utilize conditional diffusion models to learn the joint distribution of future trajectories of multiple agents, achieving multimodal and controllable trajectory prediction;
MotionTrack: Learning Robust Short-Term and Long-Term Motions for Multi-Object Tracking
Zheng Qin (Xi'an Jiaotong University), Wei Tang (University of Illinois)
Object TrackingGraph Neural NetworkVideoBenchmark
🎯 What it does: This paper proposes MotionTrack, an online multi-object tracking framework that uses an Interaction Module to handle target interactions in dense crowds and a Refind Module to re-identify long-term occluded targets through historical trajectories.
MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors
Yuang Zhang (Shanghai Jiao Tong University), Xiangyu Zhang (MEGVII Technology)
Object DetectionObject TrackingAutonomous DrivingTransformerVideo
🎯 What it does: The paper presents MOTRv2, an end-to-end multi-object tracking method that combines a pre-trained YOLOX detector with the MOTR framework.
MOVES: Manipulated Objects in Video Enable Segmentation
Richard E. L. Higgins (University of Michigan), David F. Fouhey (University of Michigan)
Object DetectionSegmentationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: By utilizing manipulation actions in videos (such as the movement of hands and objects), we learn unsupervised pixel embeddings to achieve segmentation of hands and held objects, as well as clustering of background objects.
Movies2Scenes: Using Movie Metadata To Learn Scene Representation
Shixing Chen (Amazon Prime Video), Raffay Hamid (Amazon Prime Video)
Object DetectionRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: This paper proposes using movie metadata (such as co-viewing, genre, and script summaries) for contrastive learning to train a general scene-level visual representation that supports various downstream tasks.
MP-Former: Mask-Piloted Transformer for Image Segmentation
Hao Zhang (Hong Kong University of Science and Technology), Lei Zhang (Hong Kong University of Science and Technology)
SegmentationTransformerImage
🎯 What it does: A Mask-Piloted Transformer (MP-Former) is proposed to improve the mask attention of Mask2Former, achieving more consistent multi-layer mask predictions and enhancing instance, semantic, and panoptic segmentation performance.
MSeg3D: Multi-Modal 3D Semantic Segmentation for Autonomous Driving
Jiale Li (Zhejiang University), Yong Ding (Zhejiang University)
SegmentationAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: Proposes the MSeg3D multimodal 3D semantic segmentation model, which integrates LiDAR point clouds and multi-camera information to achieve semantic prediction for full-view point clouds;
MSF: Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences
Chenhang He (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Object DetectionAutonomous DrivingComputational EfficiencyPoint CloudSequential
🎯 What it does: Proposes an efficient 3D object detection framework that generates proposals only in the current frame and propagates proposals along the time axis to sample point clouds.
MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID
Jianyang Gu (Zhejiang University), Jian Zhao (Alibaba Group)
RecognitionRetrievalDomain AdaptationNeural Architecture SearchContrastive LearningImage
🎯 What it does: A dual-contrast mechanism-based neural architecture search method is proposed for the object re-identification task, automatically designing a lightweight network called MSINet.
MSMDFusion: Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection
Yang Jiao (Fudan University), Yu-Gang Jiang (Fudan University)
Object DetectionAutonomous DrivingImagePoint Cloud
🎯 What it does: A multi-scale LiDAR-Camera fusion framework called MSMDFusion is proposed for 3D object detection, utilizing virtual points to project image semantics into 3D space.
Multi Domain Learning for Motion Magnification
Jasdeep Singh (Indian Institute of Technology Ropar), G. Sankara Raju Kosuru (Indian Institute of Technology Ropar)
Video
🎯 What it does: This paper proposes a multi-domain lightweight network that achieves video motion magnification through a combination of frequency domain phase-amplitude transformation and spatial domain multi-scale texture correction.
Multi-Agent Automated Machine Learning
Zhaozhi Wang (Peking University), Zongqing Lu (Peking University)
OptimizationHyperparameter SearchReinforcement LearningAgentic AIImage
🎯 What it does: A multi-agent automated machine learning framework MA2ML is proposed for jointly optimizing data augmentation, network architecture search, and hyperparameter optimization modules.
Multi-Centroid Task Descriptor for Dynamic Class Incremental Inference
Tenghao Cai (East China Normal University), Yuan Xie (East China Normal University)
ClassificationKnowledge DistillationImage
🎯 What it does: Proposes an independent gating network for task ID prediction in class-incremental learning, enabling dynamic inference.
Multi-Concept Customization of Text-to-Image Diffusion
Nupur Kumari (Carnegie Mellon University), Jun-Yan Zhu (Tsinghua University)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: This paper proposes an efficient fine-tuning method called Custom Diffusion, which can add new concepts to a pre-trained text-to-image diffusion model with only a few example images provided, and supports multi-concept joint generation.
Multi-Granularity Archaeological Dating of Chinese Bronze Dings Based on a Knowledge-Guided Relation Graph
Rixin Zhou (Jilin University), Chuntao Li (Jilin University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This study proposes a multi-granularity dating method for bronze tripods based on deep learning and archaeological knowledge.
Multi-Label Compound Expression Recognition: C-EXPR Database & Network
Dimitrios Kollias (Queen Mary University of London)
RecognitionTransformerVideo
🎯 What it does: The first large-scale, real-world composite expression database C-EXPR-DB is proposed, and a multi-task network C-EXPR-NET is designed to achieve composite expression recognition and action unit detection.
Multi-Level Logit Distillation
Ying Jin (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
ClassificationObject DetectionKnowledge DistillationImage
🎯 What it does: This paper proposes a multi-level aligned logit distillation method, which achieves knowledge distillation by aligning predictions at the instance, batch, and class levels, as well as enhancing predictions through temperature calibration, using only the logit outputs from the teacher model.
Multi-Modal Gait Recognition via Effective Spatial-Temporal Feature Fusion
Yufeng Cui (Beihang University), Yimei Kang (Beihang University)
RecognitionTransformerVideoMultimodality
🎯 What it does: A multi-modal gait recognition framework called MMGaitFormer based on Transformer is proposed, which integrates the spatiotemporal information of both human silhouettes and skeletons to achieve more robust gait representation.
Multi-Modal Learning With Missing Modality via Shared-Specific Feature Modelling
Hu Wang (University of Adelaide), Gustavo Carneiro (University of Surrey)
ClassificationSegmentationAuto EncoderMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper studies the problem of missing modalities in multimodal learning and proposes a shared-specific feature modeling framework called ShaSpec, which can simultaneously address various combinations of missing modalities during both training and testing phases.
Multi-Modal Representation Learning With Text-Driven Soft Masks
Jaeyoo Park (Seoul National University), Bohyung Han (Seoul National University)
RetrievalRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: A text-driven soft feature masking multimodal representation learning method is proposed, aiming to generate diverse image features through a self-supervised learning framework to improve the effectiveness of image-text matching.
Multi-Mode Online Knowledge Distillation for Self-Supervised Visual Representation Learning
Kaiyou Song (Megvii Technology), Zimeng Luo (Megvii Technology)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A multi-modal online knowledge distillation (MOKD) framework is proposed, allowing two models to enhance their visual representations through self-distillation and cross-distillation in a self-supervised learning manner.
Multi-Object Manipulation via Object-Centric Neural Scattering Functions
Stephen Tian (Stanford University), Jiajun Wu (Stanford University)
OptimizationRobotic IntelligenceGraph Neural NetworkNeural Radiance FieldImage
🎯 What it does: A low-dimensional scene representation based on Object-Centered Scattering Function (OSF) is constructed, and combined with a Graph Neural Network (GNN) dynamic model to achieve multi-object visual control; through inverse parameter estimation, both lighting and object pose are reconstructed, enabling robust manipulation under lighting changes and unknown object configurations within a Model Predictive Control (MPC) framework.
Multi-Realism Image Compression With a Conditional Generator
Eirikur Agustsson (Google Research), Fabian Mentzer (Google Research)
CompressionAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an image compression method that can control distortion and realism at the decoding end based on a real-time adjustable parameter β, allowing for the reconstruction of low distortion or high realism from the same compressed representation.
Multi-Sensor Large-Scale Dataset for Multi-View 3D Reconstruction
Oleg Voynov (Skolkovo Institute of Science and Technology), Denis Zorin (New York University)
Data SynthesisDepth EstimationImage
🎯 What it does: A large-scale dataset consisting of 107 scenes with multi-sensor data, including RGB and depth data, containing 1.4 million images, 100 viewpoints, and 14 types of lighting, equipped with high-precision SLS scanned reference geometry.
Multi-Space Neural Radiance Fields
Ze-Xin Yin (Nankai University), Bo Ren (Nankai University)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: A multi-space NeRF (MS-NeRF) method is proposed, which splits the Euclidean space into multiple virtual subspaces and uses a lightweight multi-space module to automatically handle specular reflection and refraction, addressing the multi-view consistency issue of NeRF under 360° perspectives.
Multi-View Adversarial Discriminator: Mine the Non-Causal Factors for Object Detection in Unseen Domains
Mingjun Xu (Chongqing University), Lei Zhang
Object DetectionDomain AdaptationAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A Multi-Angle Adversarial Discriminator (MAD) is proposed, which enhances target detection performance in unknown domains by mining and removing non-causal factors through frequency domain pseudo-correlation generation and multi-angle domain discrimination learning.
Multi-View Azimuth Stereo via Tangent Space Consistency
Xu Cao (Osaka University), Yasuyuki Matsushita (Osaka University)
Depth EstimationOptimizationNeural Radiance FieldImage
🎯 What it does: This paper proposes a method for reconstructing 3D shapes using only calibrated multi-view azimuth maps—Multi-View Azimuth Stereo (MVAS).
Multi-View Inverse Rendering for Large-Scale Real-World Indoor Scenes
Zhen Li (Realsee), Jiaqi Yang (Northwestern Polytechnical University)
RestorationOptimizationNeural Radiance FieldImage
🎯 What it does: This paper proposes a multi-view inverse rendering method for large-scale indoor scenes, capable of reconstructing global illumination and distinguishable SVBRDF materials from sparse HDR images.
Multi-View Reconstruction Using Signed Ray Distance Functions (SRDF)
Pierre Zins (Inria), Tony Tung (Meta Reality Labs)
Depth EstimationOptimizationPoint CloudMesh
🎯 What it does: A multi-view 3D reconstruction framework based on the Signed Ray Distance Function (SRDF) is proposed, combining depth optimization with voxel volume integration to achieve pixel-level accuracy and global consistency.
Multi-View Stereo Representation Revist: Region-Aware MVSNet
Yisu Zhang (Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies), Lixiang Lin (Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies)
RestorationSegmentationDepth EstimationConvolutional Neural NetworkPoint CloudMesh
🎯 What it does: This paper proposes a cost volume-based multi-view stereo network RA-MVSNet, which utilizes a dual-branch joint prediction of depth and signed distance field (SDF) to achieve complete reconstruction of texture-sparse areas and object boundaries.
Multiclass Confidence and Localization Calibration for Object Detection
Bimsara Pathiraja (Mohammed bin Zayed University of Artificial Intelligence), Muhammad Haris Khan (Mohammed bin Zayed University of Artificial Intelligence)
Object DetectionDomain AdaptationTransformerImage
🎯 What it does: A method for calibrating object detection models during the training phase (MCCL) is proposed, which calibrates both multi-class confidence and bounding box localization.
Multilateral Semantic Relations Modeling for Image Text Retrieval
Zheng Wang (University of Electronic Science and Technology of China), Heng Tao Shen (University of Electronic Science and Technology of China)
RetrievalGraph Neural NetworkContrastive LearningImageText
🎯 What it does: A multi-party semantic relationship modeling (MSRM) framework is proposed to address the one-to-many correspondence problem in image-text retrieval.
Multimodal Industrial Anomaly Detection via Hybrid Fusion
Yue Wang (Shanghai Jiao Tong University), Chengjie Wang (Tencent)
Anomaly DetectionTransformerContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: A multi-modal industrial defect detection framework M3DM based on RGB images and 3D point clouds is proposed.
Multimodal Prompting With Missing Modalities for Visual Recognition
Yi-Lun Lee (National Yang Ming Chiao Tung University), Chen-Yu Lee (Google)
ClassificationRecognitionTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This study addresses the issues of missing modalities and the cost of fine-tuning large models in multimodal learning, proposing a missing-aware prompts approach that enhances performance under various missing conditions by training less than 1% of the parameters on the ViLT pre-trained multimodal Transformer.
Multimodality Helps Unimodality: Cross-Modal Few-Shot Learning With Multimodal Models
Zhiqiu Lin (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
ClassificationRecognitionMeta LearningTransformerContrastive LearningImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes a cross-modal fine-tuning method that utilizes multimodal pre-trained models (such as CLIP, AudioCLIP) to map different modalities into the same representation space, treating class names and other texts as training samples to achieve cross-modal few-shot learning.
Multiple Instance Learning via Iterative Self-Paced Supervised Contrastive Learning
Kangning Liu (New York University), Carlos Fernandez-Granda (New York University)
ClassificationRepresentation LearningTransformerContrastive LearningBiomedical DataUltrasound
🎯 What it does: This paper proposes a multi-instance learning framework based on iterative self-paced supervised contrastive learning (ItS2CLR), which continuously optimizes the feature extractor through pseudo-labels, addressing the 'class collision' problem in traditional contrastive self-supervised learning under class imbalance scenarios.
Multiplicative Fourier Level of Detail
Yishun Dou (Shanghai Jiao Tong University), Bingbing Ni (Shanghai Jiao Tong University)
GenerationData SynthesisNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper proposes a new implicit neural representation framework called MFLOD, which generates linearly combinable sine basis functions by Fourier modulating local features on a multi-resolution feature grid and recursively multiplying them, thus achieving multi-level detail control.
Multiscale Tensor Decomposition and Rendering Equation Encoding for View Synthesis
Kang Han (James Cook University), Wei Xiang (La Trobe University)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: This paper proposes a Neural Radiance Feature Field (NRFF) that achieves high-quality view synthesis through multi-scale tensor decomposition and rendering equation feature encoding.
Multispectral Video Semantic Segmentation: A Benchmark Dataset and Baseline
Wei Ji (University of Alberta), Li Cheng (University of Alberta)
SegmentationContrastive LearningVideoBenchmark
🎯 What it does: This paper proposes a new task called multispectral video semantic segmentation (MVSS), constructs the MVSeg dataset containing 738 segments of synchronized RGB and thermal infrared videos, 3,545 finely annotated frames, and 26 object categories, and introduces the benchmark model MVNet.
Multivariate, Multi-Frequency and Multimodal: Rethinking Graph Neural Networks for Emotion Recognition in Conversation
Feiyu Chen (University of Electronic Science and Technology of China), Heng Tao Shen (University of Electronic Science and Technology of China)
RecognitionGraph Neural NetworkTextMultimodality
🎯 What it does: A multi-modal emotion recognition model based on graph neural networks (M3Net) is proposed for detecting emotions in dialogues.
Multiview Compressive Coding for 3D Reconstruction
Chao-Yuan Wu (Meta AI), Georgia Gkioxari (Meta AI)
GenerationDepth EstimationTransformerAuto EncoderImagePoint Cloud
🎯 What it does: A general 3D reconstruction framework called Multiview Compressive Coding (MCC) is proposed, which reconstructs complete 3D shapes by encoding features from RGB-D images and using a queryable 3D decoder to predict point cloud occupancy and color.
Music-Driven Group Choreography
Nhat Le (AIOZ), Anh Nguyen (University of Liverpool)
GenerationPose EstimationRecurrent Neural NetworkTransformerVideoMesh
🎯 What it does: This work proposes a task for generating group dance driven by audio, constructs a new large-scale dataset AIOZ-GDANCE, and presents a Transformer-based generative model GDanceR;
Mutual Information-Based Temporal Difference Learning for Human Pose Estimation in Video
Runyang Feng (Jilin University), Hyung Jin Chang (University of Birmingham)
Pose EstimationConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a framework based on temporal difference and mutual information separation (TDMI) for video human pose estimation. It extracts motion context through a multi-stage temporal difference encoder (TDE) and mines useful motion information using a representation separation module (RDM).
MV-JAR: Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training
Runsen Xu (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
Autonomous DrivingRepresentation LearningTransformerContrastive LearningPoint CloudBenchmark
🎯 What it does: A self-supervised pre-training method based on LiDAR point clouds, MV-JAR, is proposed, and a new data-efficient evaluation benchmark is established on the Waymo dataset.
MVImgNet: A Large-Scale Dataset of Multi-View Images
Xianggang Yu (Chinese University of Hong Kong, Shenzhen), Xiaoguang Han (Chinese University of Hong Kong, Shenzhen)
ClassificationObject DetectionSegmentationNeural Radiance FieldContrastive LearningImageVideoPoint Cloud
🎯 What it does: A large multi-view image dataset MVImgNet and its corresponding point cloud dataset MVPNet have been constructed and made publicly available, and their pre-training effects have been evaluated on various 2D/3D vision tasks.
N-Gram in Swin Transformers for Efficient Lightweight Image Super-Resolution
Haram Choi (Sogang University), Jihoon Yang (Sogang University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: Proposes an N-Gram context mechanism and constructs lightweight image super-resolution networks such as NGswin and SwinIR-NG based on this, addressing the issues of limited receptive fields and high computational costs of window self-attention.
NaQ: Leveraging Narrations As Queries To Supervise Episodic Memory
Santhosh Kumar Ramakrishnan (University of Texas at Austin), Kristen Grauman (Meta AI)
RetrievalRepresentation LearningTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: Automatically generate natural language queries (NaQ) using narrative texts with timestamps, and use them as additional training data to enhance the query localization model for long-term first-person videos;
NAR-Former: Neural Architecture Representation Learning Towards Holistic Attributes Prediction
Yun Yi (Xidian University), Xiaoyu Wang (Intellifusion)
Representation LearningNeural Architecture SearchTransformerNeural Radiance FieldTabular
🎯 What it does: A Transformer-based neural network architecture representation model called NAR-Former is proposed, which can encode network topology and operation information into sequences and generate a unified vector representation for predicting attributes such as model accuracy and latency.
Natural Language-Assisted Sign Language Recognition
Ronglai Zuo (Hong Kong University of Science and Technology), Brian Mak (Hong Kong University of Science and Technology)
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningVideoMultimodality
🎯 What it does: Using natural language semantics to assist sign language recognition, we propose language-aware label smoothing and cross-modal mixup techniques, and design a video-keypoint network to enhance sign language classification performance.
NeAT: Learning Neural Implicit Surfaces With Arbitrary Topologies From Multi-View Images
Xiaoxu Meng (Tencent Games), Bo Yang (Tencent Games)
GenerationData SynthesisNeural Radiance FieldImagePoint CloudMesh
🎯 What it does: The NeAT framework is proposed, utilizing multi-view images and neural implicit functions (SDF + validity branch) to achieve arbitrary topology surface reconstruction, and enabling end-to-end training through differentiable volumetric rendering.
NEF: Neural Edge Fields for 3D Parametric Curve Reconstruction From Multi-View Images
Yunfan Ye (National University of Defense Technology), Kai Xu (National University of Defense Technology)
GenerationOptimizationNeural Radiance FieldImageVideoPoint Cloud
🎯 What it does: A self-supervised Neural Implicit Field (NEF) is proposed, which is trained to obtain 3D edge density from multi-view 2D edge detection images and extracts 3D parametric curves from it.
NeFII: Inverse Rendering for Reflectance Decomposition With Near-Field Indirect Illumination
Haoqian Wu (NetEase Fuxi AI Lab), Xin Yu (The University of Queensland)
GenerationOptimizationNeural Radiance FieldImage
🎯 What it does: An end-to-end inverse rendering pipeline is proposed, utilizing Monte Carlo path tracing combined with neural radiance caching, capable of decomposing geometry, spatially varying BRDF (SVBRDF), and lighting from multi-view RGB images, particularly fine modeling of near-field indirect lighting.
Neighborhood Attention Transformer
Ali Hassani (University of Oregon), Humphrey Shi (University of Oregon)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: Proposes Neighborhood Attention (NA) and the Neighborhood Attention Transformer (NAT) based on NA, and implements an efficient CUDA/C++ extension called NATTEN.
NeMo: Learning 3D Neural Motion Fields From Multiple Video Instances of the Same Action
Kuan-Chieh Wang (Stanford University), Serena Yeung (Stanford University)
Pose EstimationOptimizationVideo
🎯 What it does: Proposes the NeMo (Neural Motion Field) framework, which jointly optimizes multiple video instances of the same action to recover accurate 3D motion and global root trajectories.
NeRDi: Single-View NeRF Synthesis With Language-Guided Diffusion As General Image Priors
Congyue Deng (Waymo), Dragomir Anguelov (Waymo)
GenerationData SynthesisDepth EstimationDiffusion modelNeural Radiance FieldImage
🎯 What it does: A 3D supervision-free, diffusion model-based single-view NeRF synthesis framework is proposed;