ICCV 2023 Papers — Page 12
IEEE/CVF International Conference on Computer Vision · 2156 papers
LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models
Chan Hee Song (Ohio State University), Yu Su (Ohio State University)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A few-shot planning framework based on large language models (LLM) called LLM-Planner is proposed, which can generate high-level sub-goal plans for embodied agents to complete complex tasks in visually perceptive environments based on natural language instructions, and can replan in real-time during execution based on perceptual information.
LMR: A Large-Scale Multi-Reference Dataset for Reference-Based Super-Resolution
Lin Zhang (University of Chinese Academy of Sciences), Zhaoxiang Zhang (University of Chinese Academy of Sciences)
RestorationSuper ResolutionGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A large-scale multi-reference super-resolution dataset LMR was constructed, and the MRefSR method was proposed to achieve joint super-resolution of multi-reference images.
LNPL-MIL: Learning from Noisy Pseudo Labels for Promoting Multiple Instance Learning in Whole Slide Image
Zhuchen Shao (Tsinghua University), Yongbing Zhang (Harbin Institute of Technology)
ClassificationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: A framework called LNPL-MIL is designed to promote multi-instance learning using noisy pseudo-labels, integrating weak labels at the WSI level (WA) and limited patch-level labels (LPA) to improve Whole-Slide Image level diagnosis and prognosis prediction.
Local and Global Logit Adjustments for Long-Tailed Learning
Yingfan Tao (Tsinghua University), Min Zheng
ClassificationConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: A multi-expert framework LGLA based on local and global Logit adjustment is proposed to address the long-tail classification problem.
Local Context-Aware Active Domain Adaptation
Tao Sun (Stony Brook University), Haibin Ling (Stony Brook University)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A locally context-aware active domain adaptation framework LADA is proposed, which selects target samples based on local inconsistencies predicted by the model and enhances model adaptation through an incremental anchor point set.
Local or Global: Selective Knowledge Assimilation for Federated Learning with Limited Labels
Yae Jee Cho (Carnegie Mellon University), Dimitrios Dimitriadis (Amazon)
Federated LearningImageBiomedical Data
🎯 What it does: Addressing the performance decline caused by the scarcity of client labels in federated learning by adaptively selecting local or global models for pseudo-labeling and training;
Localizing Moments in Long Video Via Multimodal Guidance
Wayner Barrios (Dartmouth), Bernard Ghanem (King Abdullah University of Science and Technology)
Object DetectionRetrievalTransformerVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: A two-stage multimodal guided video temporal localization framework is proposed, which first uses a guiding model to filter describable windows and then inputs them into a baseline localization model to enhance the grounding effect of long videos.
Localizing Object-Level Shape Variations with Text-to-Image Diffusion Models
Or Patashnik (Tel Aviv University), Daniel Cohen-Or (Tel Aviv University)
GenerationData SynthesisPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes a method that utilizes Prompt-Mixing and attention localization techniques to automatically generate collections of object-level shape variations without the need for additional input.
Locally Stylized Neural Radiance Fields
Hong-Wing Pang (Hong Kong University of Science and Technology), Sai-Kit Yeung (Hong Kong University of Science and Technology)
Image TranslationGenerationNeural Radiance FieldImage
🎯 What it does: Utilizing a dual-branch NeRF model combined with hash grid encoding to achieve local style transfer for 3D scenes;
Locating Noise is Halfway Denoising for Semi-Supervised Segmentation
Yan Fang (Institute of Information Science, Beijing Jiaotong University), Yunchao Wei (Institute of Information Science, Beijing Jiaotong University)
SegmentationImage
🎯 What it does: A method called Uncertainty-guided Patch CutMix (UPC) is proposed to locate and remove noise in pseudo-labels during the self-training process of semi-supervised semantic segmentation, thereby improving model performance.
Locomotion-Action-Manipulation: Synthesizing Human-Scene Interactions in Complex 3D Environments
Jiye Lee (Seoul National University), Hanbyul Joo (Seoul National University)
Data SynthesisRobotic IntelligenceReinforcement LearningAuto EncoderSequential
🎯 What it does: A unified framework called LAMA has been developed for generating natural and coherent long-term human motion in complex 3D indoor scenes, covering walking, scene interaction, and object manipulation.
LoCUS: Learning Multiscale 3D-consistent Features from Posed Images
Dominik A. Kloepfer (Visual Geometry Group University of Oxford), João F. Henriques (Australian National University)
SegmentationPose EstimationRetrievalTransformerContrastive LearningImage
🎯 What it does: Train a network to extract recognizable and 3D consistent multi-scale features, supporting landmark retrieval, localization, semantic and instance segmentation, and relative pose estimation for indoor scenes.
Logic-induced Diagnostic Reasoning for Semi-supervised Semantic Segmentation
Chen Liang (Zhejiang University), Yi Yang (Zhejiang University)
SegmentationImage
🎯 What it does: Designed and implemented LOGICDIAG, a framework that integrates symbolic logic knowledge with sub-symbolic semi-supervised semantic segmentation, using logical diagnosis to correct pseudo-labels and mitigate confirmation bias.
LogicSeg: Parsing Visual Semantics with Neural Logic Learning and Reasoning
Liulei Li (Zhejiang University), Yi Yang
SegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: LOGICSEG is proposed, a visual semantic parsing framework that combines hierarchical semantic structures with first-order logic reasoning, capable of utilizing both sub-symbolic learning and symbolic reasoning in semantic segmentation tasks.
LoGoPrompt: Synthetic Text Images Can Be Good Visual Prompts for Vision-Language Models
Cheng Shi (ShanghaiTech University), Sibei Yang (ShanghaiTech University)
ClassificationDomain AdaptationPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes the LoGoPrompt method, which utilizes synthetic text images as visual prompts and selects appropriate visual prompts through minimum-maximum contrastive learning to enhance the downstream classification performance of Vision Language Models (VLMs) such as CLIP.
LoLep: Single-View View Synthesis with Locally-Learned Planes and Self-Attention Occlusion Inference
Cong Wang (Tsinghua University), Dinesh Manocha (University of Maryland)
Image TranslationData SynthesisDepth EstimationConvolutional Neural NetworkNeural Radiance FieldImage
🎯 What it does: Achieve single-view synthesis by regressing locally learned planes from a single RGB image.
Long-Range Grouping Transformer for Multi-View 3D Reconstruction
Liying Yang (Macau University of Science and Technology), Yanyan Liang (Macau University of Science and Technology)
SegmentationGenerationTransformerPoint Cloud
🎯 What it does: This paper proposes a Long-Distance Group Attention (LGA) and an advanced upsampling decoder Transformer network (LRGT) for multi-view 3D reconstruction.
Long-range Multimodal Pretraining for Movie Understanding
Dawit Mureja Argaw, Fabian Caba Heilbron
RetrievalRepresentation LearningTransformerContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: A long-distance multimodal pre-training framework for movies has been constructed, utilizing long sequence information from three modalities: video, audio, and subtitles, to learn transferable cross-modal representations.
Long-Term Photometric Consistent Novel View Synthesis with Diffusion Models
Jason J. Yu (York University), Marcus A. Brubaker (York University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A single-view perspective extrapolation method based on latent diffusion models is proposed, which can generate multiple image sequences that are geometrically consistent with the original view under a given camera trajectory.
Look at the Neighbor: Distortion-aware Unsupervised Domain Adaptation for Panoramic Semantic Segmentation
Xu Zheng (Hong Kong University of Science and Technology), Lin Wang (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationTransformerImage
🎯 What it does: This paper proposes an unsupervised domain adaptation framework for semantic segmentation of panoramic images, utilizing neighborhood information to address the distortion issues caused by ERP projection.
Lossy and Lossless (L2) Post-training Model Size Compression
Yumeng Shi (Beihang University), Jianlei Yang (Beihang University)
ClassificationObject DetectionCompressionConvolutional Neural NetworkImage
🎯 What it does: A unified model compression framework is proposed in a post-training environment, capable of performing both lossless compression (entropy coding) and lossy compression (pruning, quantization) simultaneously.
LoTE-Animal: A Long Time-span Dataset for Endangered Animal Behavior Understanding
Dan Liu (Northwest A&F University), Jingdong Zhang (Northwest A&F University)
Object DetectionSegmentationPose EstimationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImageVideoBenchmark
🎯 What it does: We constructed and released LoTE-Animal, a multi-task dataset spanning 12 years and including 11 species of endangered animals in China (with video, images, detection/segmentation/pose/action annotations), and evaluated various mainstream computer vision models and their cross-domain generalization capabilities based on this dataset.
Low-Light Image Enhancement with Illumination-Aware Gamma Correction and Complete Image Modelling Network
Yinglong Wang (Meituan Inc.), Shuaicheng Liu (University of Electronic Science and Technology of China)
RestorationTransformerImage
🎯 What it does: A low-light image enhancement network IAGC based on illumination-aware gamma correction and complete pixel modeling Transformer is proposed.
Low-Light Image Enhancement with Multi-Stage Residue Quantization and Brightness-Aware Attention
Yunlong Liu (Xidian University), Guangming Shi (Xidian University)
RestorationAuto EncoderImage
🎯 What it does: This paper proposes a three-stage low-light image enhancement framework based on VQ-VAE, utilizing residual quantization to construct a more expressive positive light codebook, compensating for the gap between low-light features and the codebook through a query module, and further preserving detail and color consistency in the third stage by integrating branches and brightness-aware attention.
LPFF: A Portrait Dataset for Face Generators Across Large Poses
Yiqian Wu (Zhejiang University), Xiaogang Jin (Zhejiang University)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkImage
🎯 What it does: Collected 19,590 real large-pose facial images to construct the LPFF dataset, and trained 2D StyleGAN2-ada and 3D EG3D generators with it to enhance large-angle face generation and inversion effects.
LRRU: Long-short Range Recurrent Updating Networks for Depth Completion
Yufei Wang (Northwestern Polytechnical University), Yuchao Dai (Northwestern Polytechnical University)
Depth EstimationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: A lightweight deep network framework LRRU is proposed, which iteratively updates the pre-filled sparse depth map to generate a dense depth map through a learnable spatial variation kernel.
LU-NeRF: Scene and Pose Estimation by Synchronizing Local Unposed NeRFs
Zezhou Cheng (University of Massachusetts), Ameesh Makadia (Google Research)
Pose EstimationNeural Radiance FieldImageVideo
🎯 What it does: A local-to-global pipeline based on local unposed NeRF is proposed, which can estimate camera poses and neural radiance fields from a set of unposed images without relying on prior pose distributions.
Luminance-aware Color Transform for Multiple Exposure Correction
Jong-Hyeon Baek (Chungnam National University), Yeong Jun Koh (Chungnam National University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a brightness-aware color transformation (LACT) algorithm to correct overexposure and underexposure in multi-exposure images.
LVOS: A Benchmark for Long-term Video Object Segmentation
Lingyi Hong (Fudan University), Wenqiang Zhang (Fudan University)
SegmentationConvolutional Neural NetworkRecurrent Neural NetworkVideoBenchmark
🎯 What it does: A long-term video object segmentation benchmark dataset LVOS is proposed, and a model DDMemory based on a multi-scale memory bank is presented and evaluated on this dataset.
M2T: Masking Transformers Twice for Faster Decoding
Fabian Mentzer (Google Research), Michael Tschannen (Google DeepMind)
CompressionTransformerAuto EncoderImage
🎯 What it does: This paper proposes a Transformer model based on MaskGIT for neural image compression, and further introduces a dual-mask (input + attention) model M2T to significantly accelerate decoding.
MAAL: Multimodality-Aware Autoencoder-Based Affordance Learning for 3D Articulated Objects
Yuanzhi Liang (University of Technology Sydney), Yi Yang (Zhejiang University)
Robotic IntelligenceAuto EncoderMultimodalityPoint Cloud
🎯 What it does: A multi-modal autoencoder framework called MAAL is proposed for learning the manipulability of 3D articulated objects, which includes an MME encoder, action memory, and a decoder, utilizing only positive samples in an end-to-end training process.
MAGI: Multi-Annotated Explanation-Guided Learning
Yifei Zhang (Emory University), Liang Zhao (Emory University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkAuto EncoderImageBiomedical DataComputed Tomography
🎯 What it does: The study proposes a multi-annotation guided learning framework MAGI, which utilizes annotations from multiple experts to generate high-quality explanations and enhance model predictions in medical image classification.
MagicFusion: Boosting Text-to-Image Generation Performance by Fusing Diffusion Models
Jing Zhao (National University of Defense Technology), Wenjing Yang (National University of Defense Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A method based on Saliency-aware Noise Blending (SNB) is proposed, which is untrained and can directly integrate two pre-trained text-guided diffusion models during the DDIM sampling process, achieving fine-grained control over the generated images.
Make Encoder Great Again in 3D GAN Inversion through Geometry and Occlusion-Aware Encoding
Ziyang Yuan (Tsinghua University), Chun Yuan (Tsinghua University)
GenerationData SynthesisTransformerNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: Proposes a geometry-aware encoder and adaptive feature alignment based on EG3D, achieving high-quality 3D GAN inversion and editing from a single image.
Make-An-Animation: Large-Scale Text-conditional 3D Human Motion Generation
Samaneh Azadi (Meta AI), Sonal Gupta (Meta AI)
GenerationPose EstimationConvolutional Neural NetworkTransformerDiffusion modelVideoText
🎯 What it does: This paper proposes a text-conditioned 3D human motion generation framework based on diffusion models called Make-An-Animation. It first pre-trains static pose generation on a large-scale text-pseudo-pose dataset, and then fine-tunes it for dynamic motion generation by incorporating temporal convolution and attention layers.
Make-It-3D: High-fidelity 3D Creation from A Single Image with Diffusion Prior
Junshu Tang (Shanghai Jiao Tong University), Dong Chen (Microsoft Research)
GenerationData SynthesisDepth EstimationDiffusion modelNeural Radiance FieldContrastive LearningImagePoint CloudStochastic Differential Equation
🎯 What it does: A two-stage diffusion model-based method for high-fidelity 3D content generation from a single image, called Make‑It‑3D, has been designed.
MAMo: Leveraging Memory and Attention for Monocular Video Depth Estimation
Rajeev Yasarla (Qualcomm), Fatih Porikli (Qualcomm)
Depth EstimationAutonomous DrivingTransformerOptical FlowVideo
🎯 What it does: The MAMo framework is proposed, integrating memory and attention mechanisms into monocular video depth estimation, allowing any single-frame depth network to utilize temporal information during video inference to improve accuracy.
Manipulate by Seeing: Creating Manipulation Controllers from Pre-Trained Representations
Jianren Wang (Carnegie Mellon University), Abhinav Gupta (Carnegie Mellon University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningVideo
🎯 What it does: Using a pre-trained visual representation learning distance function and a dynamics model, four types of manipulation tasks (pushing, picking and placing, opening doors, and turning knobs) are trained based on low-cost human video data and implemented on robots, with direct action planning through shooting methods;
MAP: Towards Balanced Generalization of IID and OOD through Model-Agnostic Adapters
Min Zhang (Zhejiang University), Kun Kuang (Zhejiang University)
Domain AdaptationOptimizationImage
🎯 What it does: A model-agnostic method based on auxiliary adapters (MAP) is proposed, achieving a balance between IID and OOD generalization capabilities through bilevel optimization.
MAPConNet: Self-supervised 3D Pose Transfer with Mesh and Point Contrastive Learning
Jiaze Sun (Imperial College London), Tae-Kyun Kim (Korea Advanced Institute of Science and Technology)
Pose EstimationContrastive LearningPoint CloudMesh
🎯 What it does: MAPConNet is proposed, a self-supervised 3D pose transfer framework that can be trained without corresponding labels or target outputs.
MapFormer: Boosting Change Detection by Using Pre-change Information
Maximilian Bernhard (LMU Munich), Matthias Schubert (LMU Munich)
SegmentationTransformerContrastive LearningImageMultimodality
🎯 What it does: This paper proposes the use of pre-change semantic maps in remote sensing dual-phase change detection to achieve conditional change detection and cross-modal change detection;
MapPrior: Bird's-Eye View Map Layout Estimation with Generative Models
Xiyue Zhu (University of Illinois at Urbana-Champaign), Shenlong Wang (University of Illinois at Urbana-Champaign)
GenerationAutonomous DrivingTransformerGenerative Adversarial NetworkMultimodalityPoint Cloud
🎯 What it does: This paper proposes the MapPrior framework, which combines a discriminative BEV perception model with a generative map prior, enabling the generation of more accurate, realistic, and uncertainty-explainable bird's-eye view map layouts.
March in Chat: Interactive Prompting for Remote Embodied Referring Expression
Yanyuan Qiao (Australian Institute for Machine Learning), Qi Wu (Australian Institute for Machine Learning)
OptimizationRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision Language ModelText
🎯 What it does: An interactive planning framework named March-in-Chat (MiC) is proposed, enabling VLN agents to engage in real-time dialogue with large language models (LLMs) to generate fine-grained navigation plans based on high-level instructions and current visual perceptions, achieving remote entity localization.
Markov Game Video Augmentation for Action Segmentation
Nicolas Aziere (Oregon State University), Sinisa Todorovic (Oregon State University)
SegmentationConvolutional Neural NetworkReinforcement LearningVideo
🎯 What it does: In the action segmentation task, the authors propose a dual data augmentation method based on deep feature space and action sequences, aimed at improving the model's generalization ability by modifying video frame features and subtitles (action sequences).
MARS: Model-agnostic Biased Object Removal without Additional Supervision for Weakly-Supervised Semantic Segmentation
Sanghyun Jo (OGQ), Kyungsu Kim (Samsung Electronics)
Object DetectionSegmentationKnowledge DistillationImage
🎯 What it does: This paper proposes a model-agnostic bias object removal framework called MARS, which addresses the misjudgment issues in weakly supervised semantic segmentation caused by background or related objects (e.g., misclassifying a railway as a train).
MAS: Towards Resource-Efficient Federated Multiple-Task Learning
Weiming Zhuang (Sony AI), Shuai Zhang (SenseTime Research)
Federated LearningComputational EfficiencyImage
🎯 What it does: A MAS (Merge and Split) framework is proposed to address the challenge of efficiently training multi-task federated learning on resource-constrained edge devices.
MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing
Mingdeng Cao (University of Tokyo), Yinqiang Zheng (University of Tokyo)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A method called MasaCtrl is proposed, which does not require fine-tuning. By transforming self-attention in diffusion models into mutual self-attention, it enables querying the content of the source image, thereby achieving consistent image generation and complex non-rigid editing while maintaining texture and identity.
Mask-Attention-Free Transformer for 3D Instance Segmentation
Xin Lai (Chinese University of Hong Kong), Jiaya Jia (SmartMore)
Object DetectionSegmentationTransformerPoint Cloud
🎯 What it does: This paper proposes a mask-free attention Transformer framework that achieves 3D instance segmentation through an auxiliary center regression task.
Masked Autoencoders are Efficient Class Incremental Learners
Jiang-Tian Zhai (Nankai University), Ming-Ming Cheng (Nankai University)
ClassificationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: A bidirectional MAE framework based on Masked Autoencoders (MAE) is proposed for efficient category incremental learning.
Masked Autoencoders Are Stronger Knowledge Distillers
Shanshan Lao (Tsinghua University), Yujiu Yang (Tsinghua University)
Object DetectionSegmentationKnowledge DistillationTransformerAuto EncoderImage
🎯 What it does: By inputting randomly masked images into the student network, a masked autoencoder framework is used for knowledge distillation to enhance the performance of lightweight models in fine-grained visual tasks such as object detection and semantic segmentation.
Masked Diffusion Transformer is a Strong Image Synthesizer
Shanghua Gao (Nankai University), Shuicheng Yan (Sea AI Lab)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Proposes the Masked Diffusion Transformer (MDT), which enhances contextual association learning of diffusion probabilistic models through masked modeling in the latent space, significantly accelerating training and improving image generation quality.
Masked Motion Predictors are Strong 3D Action Representation Learners
Yunyao Mao (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Representation LearningTransformerVideo
🎯 What it does: A self-supervised pre-training framework MAMP based on Transformer is proposed, which learns 3D action representations by predicting corresponding motion sequences on masked skeletal sequences.
Masked Retraining Teacher-Student Framework for Domain Adaptive Object Detection
Zijing Zhao (Peking University), Yang Liu (Peking University)
Object DetectionDomain AdaptationAuto EncoderImage
🎯 What it does: A Masked Retraining Teacher-Student (MRT) method based on a teacher-student framework is proposed for unsupervised domain adaptation in object detection.
Masked Spatio-Temporal Structure Prediction for Self-supervised Learning on Point Cloud Videos
Zhiqiang Shen (Shanghai Jiao Tong University), Xi Zhou (Shanghai Jiao Tong University)
RecognitionSegmentationTransformerVideoPoint Cloud
🎯 What it does: This paper proposes a self-supervised pre-training framework called MaST-Pre, which captures spatiotemporal structures by constructing spatio-temporal tube masks on point cloud videos and jointly learning appearance reconstruction and motion prediction.
Masked Spiking Transformer
Ziqing Wang (North Carolina State University), Renjing Xu (Hong Kong University of Science and Technology)
ClassificationRecognitionComputational EfficiencyKnowledge DistillationSpiking Neural NetworkTransformerImage
🎯 What it does: A Masked Spiking Transformer (MST) model based on ANN-to-SNN conversion has been designed and implemented, utilizing Random Spiking Masking (RSM) technology to reduce redundant spikes during the inference phase, thereby lowering energy consumption.
MasQCLIP for Open-Vocabulary Universal Image Segmentation
Xin Xu (Peking University), Zhuowen Tu (University of California)
Object DetectionSegmentationKnowledge DistillationTransformerImage
🎯 What it does: A universal segmentation framework called MasQCLIP has been developed, capable of unifying instance, semantic, and panoptic segmentation tasks.
Mastering Spatial Graph Prediction of Road Networks
Anagnostidis Sotiris (ETH Zurich), Thomas Hofmann (ETH Zurich)
GenerationOptimizationConvolutional Neural NetworkTransformerReinforcement LearningImageGraph
🎯 What it does: A self-regressive graph generation framework based on reinforcement learning is proposed to predict road networks from satellite images.
MAtch, eXpand and Improve: Unsupervised Finetuning for Zero-Shot Action Recognition with Language Knowledge
Wei Lin (Graz University of Technology), Horst Bischof (Graz University of Technology)
RecognitionTransformerLarge Language ModelContrastive LearningVideoText
🎯 What it does: By unsupervisedly utilizing unlabeled videos and multi-source language knowledge (action dictionary, GPT-3 generated text, BLIP visual captions) to construct a text bag, and fine-tuning CLIP using multi-instance learning, we enhance zero-shot and few-shot action recognition performance.
MATE: Masked Autoencoders are Online 3D Test-Time Learners
M. Jehanzeb Mirza (Graz University of Technology), Horst Bischof (Graz University of Technology)
ClassificationDomain AdaptationAuto EncoderPoint Cloud
🎯 What it does: MATE is proposed, a training method for 3D point cloud testing based on a mask autoencoder, which can self-supervise adaptation on a single sample and enhance robustness against disturbances such as noise and density.
MatrixCity: A Large-scale City Dataset for City-scale Neural Rendering and Beyond
Yixuan Li (Chinese University of Hong Kong), Bo Dai (Shanghai AI Laboratory)
GenerationData SynthesisNeural Radiance FieldImageMultimodalityBenchmark
🎯 What it does: A MatrixCity dataset was constructed, and city-scale neural rendering was benchmarked on it.
MatrixVT: Efficient Multi-Camera to BEV Transformation for 3D Perception
Hongyu Zhou (MEGVII Technology), Xiangyu Zhang (MEGVII Technology)
Object DetectionDepth EstimationAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes MatrixVT, an efficient transformation method from multi-camera to BEV based on Feature Transporting Matrix.
MB-TaylorFormer: Multi-Branch Efficient Transformer Expanded by Taylor Formula for Image Dehazing
Yuwei Qiu (Sun Yat-sen University), Zhi Jin (Sun Yat-sen University)
RestorationTransformerImage
🎯 What it does: This paper proposes a multi-branch, lightweight Transformer network called MB-TaylorFormer for single image dehazing. It mainly achieves linear complexity by embedding visual tokens through multi-scale deformable convolutions and implementing Taylor expansion softmax attention, while incorporating a multi-scale attention correction module (MSAR) to compensate for Taylor approximation errors.
MBPTrack: Improving 3D Point Cloud Tracking with Memory Networks and Box Priors
Tian-Xing Xu (Tsinghua University), Song-Hai Zhang (Tsinghua University)
Object TrackingAutonomous DrivingTransformerPoint Cloud
🎯 What it does: A 3D single-object tracking framework based on memory networks, MBPTrack, is proposed, utilizing historical frames and box priors for coarse-to-fine localization.
MDCS: More Diverse Experts with Consistency Self-distillation for Long-tailed Recognition
Qihao Zhao (Beijing University of Chemical Technology), Jun Liu (Singapore University of Technology and Design)
ClassificationRecognitionKnowledge DistillationMixture of ExpertsImage
🎯 What it does: A multi-expert framework MDCS is proposed, which enhances expert diversity through diversity loss and reduces model variance via consistency self-distillation to improve long-tail classification performance.
Measuring Asymmetric Gradient Discrepancy in Parallel Continual Learning
Fan Lyu (Tianjin University), Wei Feng (Tianjin University)
ClassificationRecognitionOptimizationImage
🎯 What it does: This paper proposes and validates the use of Asymmetric Gradient Distance (AGD) and Maximum Difference Optimization (MaxDO) in Parallel Continual Learning (PCL) to alleviate the issues of gradient conflict and catastrophic forgetting.
MedKLIP: Medical Knowledge Enhanced Language-Image Pre-Training for X-ray Diagnosis
Chaoyi Wu (Cooperative Medianet Innovation Center, Shanghai Jiao Tong University), Weidi Xie (Cooperative Medianet Innovation Center, Shanghai Jiao Tong University)
ClassificationObject DetectionSegmentationTransformerContrastive LearningImageTextMultimodalityComputed Tomography
🎯 What it does: This paper proposes a medical knowledge-enhanced visual-language pre-training model MedKLIP for the diagnosis and localization of X-ray images.
MEFLUT: Unsupervised 1D Lookup Tables for Multi-exposure Image Fusion
Ting Jiang (Megvii Technology), Shuaicheng Liu (University of Electronic Science and Technology of China)
RestorationConvolutional Neural NetworkImage
🎯 What it does: An efficient method for multi-exposure image fusion using 1D LUT, called MEFLUT, is proposed.
MEGA: Multimodal Alignment Aggregation and Distillation For Cinematic Video Segmentation
Najmeh Sadoughi (Amazon Prime Video), Rohith MV (Amazon Prime Video)
SegmentationKnowledge DistillationTransformerContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: This study focuses on scene and plot segmentation in long videos (>60 minutes) and proposes MEGA (Multimodal Alignment Aggregation and Distillation), a unified multimodal Transformer to address the video segmentation problem.
Membrane Potential Batch Normalization for Spiking Neural Networks
Yufei Guo (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Intelligent Science and Technology Academy of CASIC)
Spiking Neural NetworkImage
🎯 What it does: In traditional SNNs, using only BN after convolution can disrupt membrane potential updates. The authors propose adding a layer of BN (MPBN) after the membrane potential update to maintain stable data distribution, and then merging MPBN into the threshold through reparameterization, thereby not increasing inference costs.
Memory-and-Anticipation Transformer for Online Action Understanding
Jiahao Wang (Nanjing University), Tong Lu (Nanjing University)
RecognitionTransformerOptical FlowVideo
🎯 What it does: The Memory-and-Anticipation Transformer (MAT) model is proposed, which simultaneously performs online action detection and action anticipation within a unified Transformer framework, modeling the complete temporal structure through the cyclical interaction of memory and anticipation.
MemorySeg: Online LiDAR Semantic Segmentation with a Latent Memory
Enxu Li (Waabi University of Toronto), Raquel Urtasun (Waabi University of Toronto)
SegmentationAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkPoint Cloud
🎯 What it does: Proposes MEMORYSEG, an online LiDAR semantic segmentation framework that utilizes sparse 3D implicit memory for real-time segmentation of continuous point clouds.
MeMOTR: Long-Term Memory-Augmented Transformer for Multi-Object Tracking
Ruopeng Gao (Nanjing University), Limin Wang (Nanjing University)
Object DetectionObject TrackingTransformerVideo
🎯 What it does: Proposes MeMOTR, a multi-object tracking model enhanced by long-term memory using Transformer.
Mesh2Tex: Generating Mesh Textures from Image Queries
Alexey Bokhovkin (Technical University of Munich), Angela Dai (Technical University of Munich)
GenerationData SynthesisGenerative Adversarial NetworkImageMesh
🎯 What it does: Train a hybrid mesh-neural field texture representation to learn texture manifolds and utilize this manifold for unconditional texture generation and single-image texture transfer.
Meta OOD Learning For Continuously Adaptive OOD Detection
Xinheng Wu (Australian Artificial Intelligence Institute, University of Technology Sydney), Guangquan Zhang (Australian Artificial Intelligence Institute, University of Technology Sydney)
Domain AdaptationAnomaly DetectionMeta LearningImage
🎯 What it does: A Continuous Adaptive Outlier Detection (CAOOD) framework is proposed, and a Meta OOD Learning (MOL) method is designed to quickly adapt when new distributions arrive.
Meta-ZSDETR: Zero-shot DETR with Meta-learning
Lu Zhang (Fudan University), Shuigeng Zhou (Fudan University)
Object DetectionMeta LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes Meta-ZSDETR, a zero-shot object detection framework based on Deformable DETR and meta-learning, which can directly predict bounding boxes for unseen classes during training using class-specific queries and semantic vectors, and can perform detection during testing by inputting only the semantic vectors of the unseen classes.
MetaBEV: Solving Sensor Failures for 3D Detection and Map Segmentation
Chongjian Ge (University of Hong Kong), Ping Luo (University of Hong Kong)
Object DetectionSegmentationAutonomous DrivingTransformerMixture of ExpertsMultimodalityPoint Cloud
🎯 What it does: Proposes the MetaBEV framework, which uses a BEV perspective to fuse multi-modal sensors (LiDAR and cameras) for 3D detection and BEV map segmentation, maintaining robust performance under sensor failures (complete loss or severe distortion).
MetaF2N: Blind Image Super-Resolution by Learning Efficient Model Adaptation from Faces
Zhicun Yin (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
RestorationSuper ResolutionMeta LearningConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a blind image super-resolution method based on facial information, called MetaF2N, which utilizes facial regions for model adaptation in a single low-quality image, requiring only one fine-tuning to enhance overall image quality.
MetaGCD: Learning to Continually Learn in Generalized Category Discovery
Yanan Wu (Beijing Jiaotong University), Songhe Feng (Beijing Jiaotong University)
Meta LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes a Continuous General Category Discovery (C-GCD) scenario and designs the MetaGCD method, which utilizes a meta-learning framework to train in the offline phase for the unsupervised discovery of new categories during the online incremental phase while maintaining performance on known categories.
Metric3D: Towards Zero-shot Metric 3D Prediction from A Single Image
Wei Yin (DJI Technology), Chunhua Shen
Depth EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: A zero-shot single-view metric depth prediction framework based on 'Standard Camera Space Transformation' (CSTM) and random local normalization loss is proposed, which can directly recover the real-scale 3D scene from a single image.
MeViS: A Large-scale Benchmark for Video Segmentation with Motion Expressions
Henghui Ding (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
SegmentationTransformerVideoTextBenchmark
🎯 What it does: This study first constructs a large-scale video segmentation dataset, MeViS, focusing on guiding target object segmentation through motion expression, and proposes a baseline method, LMPM, based on language queries and motion perception.
MGMAE: Motion Guided Masking for Video Masked Autoencoding
Bingkun Huang (Nanjing University), Limin Wang (Nanjing University)
Representation LearningTransformerAuto EncoderOptical FlowVideo
🎯 What it does: This paper proposes the Motion Guided Masked Autoencoder (MGMAE), which utilizes a time-consistent masking strategy guided by optical flow to dynamically select visible tokens, enhancing the effectiveness of self-supervised pre-training for videos.
MHCN: A Hyperbolic Neural Network Model for Multi-view Hierarchical Clustering
Fangfei Lin (University of Electronic Science and Technology of China), Zenglin Xu (University of Electronic Science and Technology of China)
Representation LearningAuto EncoderImage
🎯 What it does: This paper proposes a one-time training multi-view hierarchical clustering network (MHCN) that utilizes a multi-view hyper-surface autoencoder to learn a unified hierarchical embedding on the Poincaré sphere. It achieves unified optimization of multi-view consistency, view specificity, and tree balance through a triple loss of alignment, reconstruction, and uniformity.
MHEntropy: Entropy Meets Multiple Hypotheses for Pose and Shape Recovery
Rongyu Chen (National University of Singapore), Angela Yao (National University of Singapore)
Pose EstimationConvolutional Neural NetworkFlow-based ModelImage
🎯 What it does: A multi-hypothesis probabilistic framework is proposed to estimate 3D pose and shape from a monocular RGB image, generating diverse and physically constrained solutions in occluded and ambiguous scenes.
MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices
Andranik Sargsyan (Picsart AI Research), Humphrey Shi (Georgia Institute of Technology)
RestorationGenerationKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A lightweight mobile image inpainting network MI-GAN is proposed, which significantly reduces the number of parameters and computational power while maintaining high quality.
Mimic3D: Thriving 3D-Aware GANs via 3D-to-2D Imitation
Xingyu Chen, Baoyuan Wang
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A 3D-to-2D imitation learning strategy is proposed, allowing the direct 3D rendering branch to produce high-quality images while maintaining strict 3D consistency, and incorporating 3D-aware convolution in the three-plane generator to enhance feature representation.
MIMO-NeRF: Fast Neural Rendering with Multi-input Multi-output Neural Radiance Fields
Takuhiro Kaneko (NTT Corporation)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: This paper proposes MIMO-NeRF, which replaces the single-input single-output MLP of NeRF with a multi-input multi-output MLP, and utilizes self-supervised learning to address the ambiguity of color and volume density, thereby accelerating rendering.
Minimal Solutions to Generalized Three-View Relative Pose Problem
Yaqing Ding (Lund University), Benjamin Kimia (Lund University)
Pose EstimationSimultaneous Localization and MappingImage
🎯 What it does: This paper addresses the non-central (generalized) camera model and proposes two methods for solving the minimum relative pose problem under three views: the four-point method and the six-line method.
Minimal Solutions to Uncalibrated Two-view Geometry with Known Epipoles
Gaku Nakano (NEC Corporation)
Pose EstimationOptimizationComputational EfficiencyImageVideo
🎯 What it does: This paper proposes four minimal solutions for uncalibrated dual-view geometry using known pole information (2 points + f, 4 points + f, 3 points + f + k, 5 points + f + k), provides a closed-form univariate polynomial root-finding method, and incorporates pole regularization into RANSAC + nonlinear refinement.
Minimum Latency Deep Online Video Stabilization
Zhuofan Zhang (University of Electronic Science and Technology of China), Shuaicheng Liu (University of Electronic Science and Technology of China)
RestorationData SynthesisConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: A framework for online video stabilization is proposed, which first estimates camera motion using deep mesh flow, then smooths the motion trajectory with a network, and finally generates a stable view from the last frame.
Mining bias-target Alignment from Voronoi Cells
Rémi Nahon (Telecom Paris), Enzo Tartaglione (Telecom Paris)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: An unsupervised, bias-agnostic debiasing method is proposed, which identifies bias-target alignment information by extracting the Voronoi boundary distance of samples at the bottleneck layer, and uses this information to weight and regularize the training process, thereby reducing the deep network's dependence on bias in the data.
MiniROAD: Minimal RNN Framework for Online Action Detection
Joungbin An (Yonsei University), Seon Joo Kim (Yonsei University)
RecognitionAnomaly DetectionComputational EfficiencyRecurrent Neural NetworkOptical FlowVideoSequential
🎯 What it does: This paper proposes a lightweight RNN framework called MiniROAD for online action detection, addressing the mismatch between training and inference phases;
Misalign, Contrast then Distill: Rethinking Misalignments in Language-Image Pre-training
Bumsoo Kim (LG AI Research), Seunghwan Kim (LG AI Research)
RetrievalKnowledge DistillationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper addresses the issue of text-image misalignment caused by random image augmentation in language-image pretraining, proposing a framework (MCD) that utilizes misalignment information for training.
Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning
Byung-Kwan Lee (Korea Advanced Institute of Science and Technology), Yong Man Ro (Korea Advanced Institute of Science and Technology)
ClassificationAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: A framework named Adversarial Double Machine Learning (ADML) is proposed, which utilizes causal inference to estimate the causal effect of adversarial perturbations on model predictions, and based on this, reduces the effect through reweighting training to enhance the model's robustness against adversarial attacks.
Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground
Haoxin Li (Nanyang Technological University), Boyang Li (Nanyang Technological University)
RecognitionDomain AdaptationConvolutional Neural NetworkVideoBenchmark
🎯 What it does: This paper proposes the StillMix method, which utilizes a 2D reference network to identify and mix static frames to suppress background and foreground static biases in video action recognition.
MixBag: Bag-Level Data Augmentation for Learning from Label Proportions
Takanori Asanomi (Kyushu University), Ryoma Bise (Kyushu University)
ClassificationData-Centric LearningImage
🎯 What it does: A data augmentation method based on bags, MixBag, and the corresponding confidence interval loss are proposed to address the issue of accuracy decline due to the lack of bag quantity in Learning with Label Proportions (LLP);
MixCycle: Mixup Assisted Semi-Supervised 3D Single Object Tracking with Cycle Consistency
Qiao Wu (Northwestern Polytechnical University), Mathieu Salzmann (École Polytechnique Fédérale de Lausanne)
Object TrackingAutonomous DrivingPoint Cloud
🎯 What it does: MixCycle is proposed, a semi-supervised 3D single object tracking framework that utilizes self-tracking loops, forward and backward loops, and SOTMixup data augmentation specifically designed for SOT.
Mixed Neural Voxels for Fast Multi-view Video Synthesis
Feng Wang (Tsinghua University), Huaping Liu (Tsinghua University)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldVideo
🎯 What it does: By using a mixed voxel representation, the 4D dynamic scene is split into static and dynamic voxels, which are rendered using lightweight and dynamic networks respectively, achieving fast training and high frame rate rendering.
MixPath: A Unified Approach for One-shot Neural Architecture Search
Xiangxiang Chu, Bo Zhang
ClassificationNeural Architecture SearchImage
🎯 What it does: A multi-path one-shot neural network architecture search framework called MixPath is proposed, which addresses the issues of unstable training and poor ranking ability of multi-path super networks by introducing Shadow Batch Normalization (SBN).
MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation
Kaixin Cai (University of Technology Sydney), Xiaodan Liang (University of Technology Sydney)
SegmentationTransformerContrastive LearningImageText
🎯 What it does: A MixReorg pre-training framework is proposed, which achieves fine-grained patch-text correspondence through mixed patch reorganization of image-text pairs, thereby learning masks for open-world semantic segmentation.
MixSpeech: Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition
Xize Cheng (Zhejiang University), Zhou Zhao (Zhejiang University)
RecognitionImage TranslationTransformerVideoMultimodalityAudio
🎯 What it does: This paper proposes the first baseline for cross-lingual visual speech translation (lip reading translation) and constructs the AVMuST-TED dataset. It introduces the MixSpeech framework, which achieves audio-visual cross-modal knowledge transfer through mixed speech self-learning.