CVPR 2025 Papers — Page 15
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
Learning Compatible Multi-Prize Subnetworks for Asymmetric Retrieval
Yushuai Sun (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
RetrievalOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: A prunable network (PrunNet) is proposed, which learns compatible sub-networks with variable capacities within a single dense network, enabling the generation of lightweight models of any capacity for multi-platform deployment without the need for retraining.
Learning Conditional Space-Time Prompt Distributions for Video Class-Incremental Learning
Xiaohan Zou (Pennsylvania State University), Shu Zhao (Pennsylvania State University)
ClassificationRecognitionTransformerPrompt EngineeringDiffusion modelVideo
🎯 What it does: CoSTEP achieves incremental learning of video categories by learning the spatial-temporal prompt distribution of videos, avoiding the limitations of traditional fixed prompt pools.
Learning Dynamic Collaborative Network for Semi-supervised 3D Vessel Segmentation
Jiao Xu (Dalian University of Technology), Lihe Zhang (Dalian University of Technology)
SegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageBiomedical Data
🎯 What it does: A dynamic collaborative network named DiCo is proposed for semi-supervised 3D vascular segmentation.
Learning Endogenous Attention for Incremental Object Detection
Xiang Song (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
Object DetectionTransformerImage
🎯 What it does: This paper proposes an incremental object detection method LEA based on an internal attention mechanism, aimed at addressing the issue of incomplete annotations in incremental detection.
Learning Extremely High Density Crowds as Active Matters
Feixiang He (University College London), He Wang (University College London)
Autonomous DrivingExplainability and InterpretabilityOptical FlowVideoPhysics RelatedStochastic Differential Equation
🎯 What it does: Research on motion prediction and analysis of extremely high-density crowds.
Learning Flow Fields in Attention for Controllable Person Image Generation
Zijian Zhou (Meta AI), Sen He (Meta AI)
GenerationData SynthesisPose EstimationConvolutional Neural NetworkDiffusion modelAuto EncoderImage
🎯 What it does: A regularization loss for learning flow fields in attention (Leffa) is proposed to control the appearance and pose in portrait generation, significantly reducing detail distortion.
Learning from Neighbors: Category Extrapolation for Long-Tail Learning
Shizhen Zhao (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningImage
🎯 What it does: Utilize GPT-4 to automatically retrieve fine-grained open-set categories related to the target class, scrape and filter web images, and construct auxiliary data; train the original data using neighbor elimination loss in long-tail classification, and directly mask auxiliary category weights during inference.
Learning from Streaming Video with Orthogonal Gradients
Tengda Han (Google DeepMind), Andrew Zisserman (University of Oxford)
OptimizationRepresentation LearningVideo
🎯 What it does: This paper proposes to enhance representation learning in continuous video stream self-supervised learning by using orthogonal gradient projection in the optimizer to reduce gradient correlation.
Learning from Synchronization: Self-Supervised Uncalibrated Multi-View Person Association in Challenging Scenes
Keqi Chen (University of Strasbourg), Nicolas Padoy (University of Strasbourg)
RecognitionObject DetectionObject TrackingContrastive LearningImageVideo
🎯 What it does: A label-free self-supervised multi-view human association method called Self-MVA is proposed, which utilizes view synchronization information to learn geometric and appearance features, achieving cross-view human association in similar appearance scenes.
Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing
Ruiyi Wang (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
RestorationGenerationDiffusion modelImage
🎯 What it does: A mist removal pipeline based on diffusion models is proposed, which first generates high-quality real fog images using HazeGen, and then performs dehazing using DiffDehaze.
Learning Heterogeneous Tissues with Mixture of Experts for Gigapixel Whole Slide Images
Junxian Wu (Southeast University), Youyong Kong (Southeast University)
ClassificationSegmentationGraph Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: This paper proposes a pluggable Pathology-Aware Mixture of Experts (PAMoE) module to capture tissue heterogeneity and perform survival prediction on whole slide images (WSI).
Learning Occlusion-Robust Vision Transformers for Real-Time UAV Tracking
You Wu (Guilin University of Technology), Shuiwang Li (Guilin University of Technology)
Object TrackingComputational EfficiencyKnowledge DistillationTransformerVideo
🎯 What it does: To address the occlusion problem in drone tracking, this paper proposes the ORTrack framework based on Vision Transformer, utilizing spatial Cox process random occlusion to achieve occlusion-robust feature learning, and generating a more efficient student model through adaptive feature knowledge distillation;
Learning on Model Weights using Tree Experts
Eliahu Horwitz (Hebrew University of Jerusalem), Yedid Hoshen (Hebrew University of Jerusalem)
ClassificationRepresentation LearningMixture of ExpertsContrastive LearningImage
🎯 What it does: This paper studies how to infer the categories and functions of training data solely based on model weights, proposing the training of a lightweight probe expert ProbeX within the model tree and achieving a mapping from weights to text embeddings.
Learning Partonomic 3D Reconstruction from Image Collections
Xiaoqian Ruan (University of Illinois Chicago), Wei Tang (University of Illinois Chicago)
SegmentationGenerationTransformerNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes a hierarchical structured 3D reconstruction method based on a single-view image set, capable of simultaneously recovering overall shape and semantic parts under the condition of having only 2D mask labels.
Learning Person-Specific Animatable Face Models from In-the-Wild Images via a Shared Base Model
Yuxiang Mao (Institute of Computing Technology Chinese Academy of Sciences), Shihong Xia (Institute of Computing Technology Chinese Academy of Sciences)
RestorationGenerationTransformerAuto EncoderImage
🎯 What it does: This paper proposes a method that first trains a large-scale self-supervised ViT-MAE base model, and then personalizes the transfer by inserting lightweight adapters into the Transformer layers, achieving 3D facial reconstruction and animation for single or multiple images.
Learning Phase Distortion with Selective State Space Models for Video Turbulence Mitigation
Xingguang Zhang (Purdue University), Stanley H. Chan (Purdue University)
RestorationAuto EncoderVideo
🎯 What it does: This study investigates the problem of video turbulence attenuation and proposes the network MambaTM for simultaneous estimation and removal of turbulence.
Learning Physics From Video: Unsupervised Physical Parameter Estimation for Continuous Dynamical Systems
Alejandro Castañeda Garcia (Delft University of Technology), Nergis Tomen
OptimizationComputational EfficiencyRepresentation LearningRobotic IntelligenceVideoPhysics RelatedOrdinary Differential Equation
🎯 What it does: An unsupervised method is proposed to estimate physical parameters from videos of single-segment continuous dynamical systems. An encoder is used to map frames to latent space, and a physical block predicts the next latent state based on known differential equations, avoiding model collapse through KL divergence loss, without the need to reconstruct a decoder.
Learning Physics-Based Full-Body Human Reaching and Grasping from Brief Walking References
Yitang Li (Tsinghua University), Li Yi (Tsinghua University)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkVideoPhysics Related
🎯 What it does: This work proposes a framework that generates diverse and physically feasible full-body human reaching and grasping actions using only short-term walking MoCap data.
Learning Temporally Consistent Video Depth from Video Diffusion Priors
Jiahao Shao (Zhejiang University), Yiyi Liao (Zhejiang University)
GenerationDepth EstimationDiffusion modelAuto EncoderVideo
🎯 What it does: A self-regressive depth estimation method called ChronoDepth is proposed, which utilizes video diffusion models to generate depth maps with high spatial accuracy and temporal consistency for videos of any length.
Learning Textual Prompts for Open-World Semi-Supervised Learning
Yuxin Fan (Shanxi University), Jiye Liang (Shanxi University)
ClassificationRecognitionTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes a framework for open-world semi-supervised learning, TP-OWSSL, which combines global and local text prompt learning with a forward-backward strategy to enhance the ability to recognize known and unknown categories in unlabeled samples.
Learning to Detect Objects from Multi-Agent LiDAR Scans without Manual Labels
Qiming Xia (Xiamen University), Chenglu Wen (Xiamen University)
Object DetectionAutonomous DrivingContrastive LearningPoint Cloud
🎯 What it does: This paper proposes an unsupervised multi-agent LiDAR 3D object detection framework called DOtA, which utilizes shared pose and shape information to train an initial detector and enhances the quality of pseudo-labels through multi-scale boundary encoding and label-internal contrastive learning, ultimately training the detector.
Learning to Filter Outlier Edges in Global SfM
Nicole Damblon (ETH Zurich), Daniel Barath (HUN-REN SZTAKI)
Pose EstimationAnomaly DetectionGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes a Transformer-based graph neural network for global SfM, which performs binary classification of relative displacement edges by transforming the edges of the view graph into vertices of a line graph, thereby filtering out outlier edges to improve camera pose estimation.
Learning to Highlight Audio by Watching Movies
Chao Huang (University of Rochester), Sanjeel Parekh (University of Rochester)
GenerationData SynthesisTransformerVideoMultimodalityAudio
🎯 What it does: This paper proposes a vision-guided audio highlighting task, aiming to transform rough mixed audio into high-quality audio that matches the visual content of the video and maintains audio layer balance.
Learning to Normalize on the SPD Manifold under Bures-Wasserstein Geometry
Rui Wang (Jiangnan University), Xiao-Jun Wu (Jiangnan University)
ClassificationRecognitionConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This paper proposes a General Bures-Wasserstein metric (GBW)-based SPD (Symmetric Positive Definite Matrix) batch normalization (GBWBN), and introduces learnable SPD parameters and power transformations to achieve adaptive normalization of batch features.
Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation
Taeyoung Yun (KAIST), Ling Pan (Hong Kong University of Science and Technology)
GenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelImageText
🎯 What it does: This study proposes the use of GFlowNet for prompt adaptation, learning to generate prompts that are both high-quality and diverse in text-to-image generation, addressing the mode collapse issue caused by traditional RL methods.
Learning Visual Composition through Improved Semantic Guidance
Austin Stone (Google), Jonathon Shlens (Google)
RetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper enhances the performance of visual-language models in visual composition understanding and retrieval tasks by regenerating semantically rich long text descriptions from large-scale image data and fine-tuning the CLIP model with a powerful pre-trained text encoder.
Learning Visual Generative Priors without Text
Shuailei Ma (Northeastern University), Yujun Shen (Ant Group)
Image TranslationGenerationData SynthesisTransformerDiffusion modelContrastive LearningImageVideoText
🎯 What it does: A purely visual self-supervised image generation framework called Lumos is constructed, which first trains an image-to-image (I2I) model using a large number of unlabeled images, and then transfers it as a prior to downstream generation tasks such as text-to-image, viewpoint synthesis, and image-to-video.
Learning with Noisy Triplet Correspondence for Composed Image Retrieval
Shuxian Li (Sichuan University), Peng Hu (Sichuan University)
RetrievalVision Language ModelContrastive LearningImage
🎯 What it does: This paper addresses the issue of noise in triplets caused by manual annotation (NTC) and proposes a robust learning framework named TME, which can achieve high-quality composite image retrieval in noisy environments.
Learning-enabled Polynomial Lyapunov Function Synthesis via High-Accuracy Counterexample-Guided Framework
Hanrui Zhao (East China Normal University), Zhengfeng Yang (East China Normal University)
OptimizationSupervised Fine-TuningTime Series
🎯 What it does: A learning-based polynomial Lyapunov function synthesis framework, SynNLF, is proposed, which combines polynomial neural network training with SOS verification and accelerates the learning process through high-precision counterexample generation.
LEDiff: Latent Exposure Diffusion for HDR Generation
Chao Wang (MPI Informatik), Xuaner Zhang (Adobe)
RestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: The LEDiff method is proposed, utilizing a pre-trained latent diffusion model to achieve high dynamic range (HDR) image generation and LDR to HDR conversion in the latent space, while maintaining the generative capabilities of the original model.
LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging
Maximilian Rokuss (German Cancer Research Center), Klaus Maier-Hein (German Cancer Research Center)
Object TrackingSegmentationConvolutional Neural NetworkPrompt EngineeringImageBiomedical DataComputed Tomography
🎯 What it does: We propose LesionLocator, an end-to-end framework capable of zero-shot, global tumor segmentation and longitudinal tracking in three-dimensional whole-body medical imaging.
Less Attention is More: Prompt Transformer for Generalized Category Discovery
Wei Zhang (Beijing Jiaotong University), Jianping Fan (Lenovo)
ClassificationRecognitionTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: Proposes the AptGCD framework, which integrates Meta Visual Prompt and Prompt Transformer to achieve universal category discovery.
Less is More: Efficient Image Vectorization with Adaptive Parameterization
Kaibo Zhao (Xidian University), Xiaotian Qiao (Xidian University)
Image TranslationComputational EfficiencyDiffusion modelImage
🎯 What it does: Proposes AdaVec, an efficient image vectorization method with adaptive path and control point numbers.
Less is More: Efficient Model Merging with Binary Task Switch
Biqing Qi (Shanghai Artificial Intelligence Laboratory), Bowen Zhou (Tsinghua University)
ClassificationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: Proposed binary task switchers (T-Switch and Auto-Switch) for efficient fusion of multiple fine-tuned models, significantly reducing task vector storage and enhancing fusion performance.
Lessons and Insights from a Unifying Study of Parameter-Efficient Fine-Tuning (PEFT) in Visual Recognition
Zheda Mai (Ohio State University), Wei-Lun Chao (Ohio State University)
RecognitionDomain AdaptationTransformerSupervised Fine-TuningImage
🎯 What it does: Systematically evaluated and compared various parameter-efficient fine-tuning (PEFT) methods for visual Transformers, conducting unified experiments on their performance in low-sample, sufficient-sample, and distribution shift scenarios.
Let Humanoids Hike! Integrative Skill Development on Complex Trails
Kwan-Yee Lin (University of Michigan), Stella X. Yu (University of Michigan)
Robotic IntelligenceTransformerReinforcement LearningAgentic AIMultimodality
🎯 What it does: An end-to-end framework named LEGO-H has been developed and trained, enabling humanoid robots to autonomously walk on complex paths and complete tasks.
Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples
Weiwei Li (University of Electronic Science and Technology of China), Wen Li (University of Electronic Science and Technology of China)
ClassificationData-Centric LearningSupervised Fine-TuningImageText
🎯 What it does: A data-driven debiasing method that does not require biased labels is proposed, which detects and neutralizes biased features by identifying the dispersion of samples in the feature space, thereby improving the worst group accuracy of the model.
Let's Chorus: Partner-aware Hybrid Song-Driven 3D Head Animation
Xiumei Xie (South China University of Technology), Huaidong Zhang (South China University of Technology)
GenerationPose EstimationTransformerAuto EncoderVideoAudio
🎯 What it does: Proposes the PaChorus framework, which implements multi-singer remix-driven 3D head animation;
Let's Verify and Reinforce Image Generation Step by Step
Renrui Zhang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationReinforcement LearningImageTextChain-of-Thought
🎯 What it does: This paper transfers the Chain of Thought (CoT) strategy to autoregressive image generation tasks and improves generation quality through methods such as testing-time validation, preference alignment, and a combination of both.
Leveraging 3D Geometric Priors in 2D Rotation Symmetry Detection
Ahyun Seo (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
Object DetectionSegmentationTransformerImage
🎯 What it does: A rotation symmetry detection framework based on 3D geometric priors is proposed, which directly predicts the rotation center and seed vertices in the 3D camera coordinate system. It then reconstructs the complete vertex set through the rotation axis and symmetry order, and projects it back to the 2D image.
Leveraging Global Stereo Consistency for Category-Level Shape and 6D Pose Estimation from Stereo Images
Junning Qiu (Xi'an Jiaotong University), Yonggen Ling (Tencent Robotics X)
Pose EstimationRobotic IntelligenceImage
🎯 What it does: A method for category-level shape and 6D pose estimation based on binocular images using global stereo consistency constraints is proposed.
Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection
Wenxi Chen (Purdue University), Yan Gu (Purdue University)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a post-processing method called PRO, which minimizes the softmax score by performing a limited perturbation search on the input to enhance OOD detection.
Leveraging SD Map to Augment HD Map-based Trajectory Prediction
Zhiwei Dong (Huawei Technologies), Jia Guo (Huawei Technologies)
Autonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: The SATP framework is proposed, which integrates Standard Definition maps (SD maps) with online High Definition maps (HD maps) for trajectory prediction.
Leveraging Temporal Cues for Semi-Supervised Multi-View 3D Object Detection
Jinhyung Park (Carnegie Mellon University), Kris Kitani (Carnegie Mellon University)
Object DetectionAutonomous DrivingAuto EncoderVideo
🎯 What it does: A semi-supervised 3D object detection framework for temporal RGB videos is proposed, which enhances pseudo-label quality and detection performance through forward and backward temporal training, 2D-3D Hungarian matching, tracking completion, and object query-based occlusion reconstruction.
LeviTor: 3D Trajectory Oriented Image-to-Video Synthesis
Hanlin Wang (Nanjing University), Limin Wang (Nanjing University)
SegmentationGenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: The LeviTor method is proposed, achieving image-to-video synthesis based on 3D trajectory control.
Libra-Merging: Importance-redundancy and Pruning-merging Trade-off for Acceleration Plug-in in Large Vision-Language Model
Longrong Yang (Zhejiang University), Xi Li (Kuaishou Technology)
CompressionComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: A training-independent visual token compression plugin called Libra-Merging is designed for accelerating inference in large vision-language models.
LibraGrad: Balancing Gradient Flow for Universally Better Vision Transformer Attributions
Faridoun Mehri (Sharif University of Technology), Mohammad Taher Pilehvar (Cardiff University)
Explainability and InterpretabilityTransformerImage
🎯 What it does: This paper proposes LibraGrad, a post-hoc explanation method that balances gradient flow by pruning and scaling the backward paths of Transformers, enhancing the reliability of gradient attribution.
LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation
Chenxu Zhou (Zhejiang University), Xiaowei Zhou (Zhejiang University)
Autonomous DrivingComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: A LiDAR-RT framework is proposed, achieving real-time, physically accurate LiDAR re-simulation in dynamic driving scenarios.
LidarGait++: Learning Local Features and Size Awareness from LiDAR Point Clouds for 3D Gait Recognition
Chuanfu Shen (University of Electronic Science and Technology of China), Shiqi Yu (Southern University of Science and Technology)
RecognitionPoint Cloud
🎯 What it does: This paper studies 3D gait recognition based on LiDAR point clouds and proposes the LidarGait++ framework, aiming to directly learn gait features from unordered point clouds.
Lifelong Knowledge Editing for Vision Language Models with Low-Rank Mixture-of-Experts
Qizhou Chen (East China Normal University), Xiaofeng He (East China Normal University)
TransformerLarge Language ModelMixture of ExpertsVision Language ModelContrastive LearningTextMultimodality
🎯 What it does: This paper proposes LiveEdit, a complete framework for lifelong knowledge editing targeting multimodal large language models (VLLMs).
Lift3D Policy: Lifting 2D Foundation Models for Robust 3D Robotic Manipulation
Yueru Jia (Peking University), Shanghang Zhang (Peking University)
Depth EstimationKnowledge DistillationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerContrastive LearningVideoMultimodalityPoint Cloud
🎯 What it does: This paper presents Lift3D, which can elevate 2D vision foundation models (such as CLIP, DINOV2) to 3D robotic manipulation. It learns depth information through task-aware MAE and directly inputs point clouds into a 2D Transformer via 2D positional embedding mapping, forming a complete 3D control strategy.
Lifting Motion to the 3D World via 2D Diffusion
Jiaman Li (Stanford University), Jiajun Wu (Stanford University)
Pose EstimationOptimizationTransformerDiffusion modelScore-based ModelVideo
🎯 What it does: The MVLift framework is proposed, which can recover global 3D motion (joint rotation + root trajectory) using only single-view 2D joint sequences.
Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference
Hao Yin (University of Science and Technology of China), Zilei Wang (University of Science and Technology of China)
OptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper explores the flow mechanism of visual information at different network depths in multimodal large language models (MLLMs) and proposes a hierarchical visual token pruning method, HiMAP, to achieve inference acceleration.
Light Transport-aware Diffusion Posterior Sampling for Single-View Reconstruction of 3D Volumes
Ludwic Leonard (Technical University of Munich), Rüdiger Westermann (Technical University of Munich)
RestorationGenerationDiffusion modelPoint Cloud
🎯 What it does: A single-view cloud volume reconstruction method based on the combination of unconditional diffusion models and differentiable volume renderers is proposed.
Light3R-SfM: Towards Feed-forward Structure-from-Motion
Sven Elflein (NVIDIA), Laura Leal-Taixé (NVIDIA)
Pose EstimationAutonomous DrivingComputational EfficiencyTransformerSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: We propose Light3R-SfM, a fully forward-learning structure from motion (SfM) framework that can directly predict globally aligned camera poses and dense point clouds on large-scale image collections without the need for traditional matching and global optimization.
LightLoc: Learning Outdoor LiDAR Localization at Light Speed
Wen Li (Xiamen University), Cheng Wang (Xiamen University)
Pose EstimationOptimizationComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: LightLoc is proposed, a new method that can quickly learn large outdoor LiDAR localization within 1 hour.
LIM: Large Interpolator Model for Dynamic Reconstruction
Remy Sabathier (University College London), David Novotny (Meta)
GenerationData SynthesisTransformerDiffusion modelMesh
🎯 What it does: A Large Interpolator Model (LIM) based on Transformer is proposed, which can perform continuous interpolation of 3D implicit representations between two key time frames, outputting high-quality 4D dynamic reconstructions.
LiMoE: Mixture of LiDAR Representation Learners from Automotive Scenes
Xiang Xu (Nanjing University of Aeronautics and Astronautics), Qingshan Liu (Nanjing University of Posts and Telecommunications)
SegmentationAutonomous DrivingKnowledge DistillationRepresentation LearningMixture of ExpertsContrastive LearningPoint Cloud
🎯 What it does: This paper proposes the LiMoE framework, which integrates three types of LiDAR representations (range images, sparse voxels, and raw point clouds) through Mixture of Experts (MoE) to achieve pre-training and downstream segmentation tasks.
Linear Attention Modeling for Learned Image Compression
Donghui Feng (Shanghai Jiao Tong University), Li Song (Shanghai Jiao Tong University)
CompressionConvolutional Neural NetworkImage
🎯 What it does: A linear attention-based image compression framework LALIC is proposed, utilizing Bi-RWKV blocks for efficient feature extraction and constructing RWKV-SCCTX as the entropy model.
LineArt: A Knowledge-guided Training-free High-quality Appearance Transfer for Design Drawing with Diffusion Model
Xi Wang (Jilin University), Chuntao Li (Jilin University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper presents LineArt, a high-quality appearance transfer framework that requires no training, capable of accurately transferring the material features of reference images to professional design drawings while maintaining the integrity of details and structure.
LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity
Hongjie Wang (Princeton University), Xiaoliang Dai (Meta)
GenerationData SynthesisComputational EfficiencyDiffusion modelVideoText
🎯 What it does: The LinGen framework is proposed, achieving linear complexity in text-to-video generation, capable of generating high-resolution videos of minutes in length on a single GPU.
Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition
Yifei Zhang (Institute of Information Engineering, Chinese Academy of Sciences), Xiangyang Ji (Tsinghua University)
RecognitionTransformerContrastive LearningImageText
🎯 What it does: A self-supervised text recognition pre-training method called LMIM is proposed, which achieves the fusion of visual and language information by incorporating language guidance during the masked image reconstruction process.
Link to the Past: Temporal Propagation for Fast 3D Human Reconstruction from Monocular Video
Matthew Marchellus (Inha University), In Kyu Park (Inha University)
Pose EstimationComputational EfficiencyVideo
🎯 What it does: The TemPoFast3D method is proposed to achieve fast 3D human reconstruction based on monocular video, utilizing temporal consistency to propagate geometric information across consecutive frames.
Link-based Contrastive Learning for One-Shot Unsupervised Domain Adaptation
Yue Zhang (Wuhan University), Chao Liang (Wuhan University)
Domain AdaptationConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: In the extreme one-shot unsupervised domain adaptation (OSUDA) scenario, a Link Contrastive Learning (LCL) method is proposed, which utilizes clustering links within the target domain to learn discriminative features and achieves semantic alignment through bidirectional matching of cross-domain links.
LION-FS: Fast & Slow Video-Language Thinker as Online Video Assistant
Wei Li (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
RecognitionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: A real-time online assistant LION-FS based on fast and slow paths is proposed, capable of processing high frame rate first-person videos and generating accurate responses.
LIRM: Large Inverse Rendering Model for Progressive Reconstruction of Shape, Materials and View-dependent Radiance Fields
Zhengqin Li (Meta Reality Labs), Zhao Dong (Meta Reality Labs)
TransformerNeural Radiance FieldImage
🎯 What it does: This paper proposes a large-scale inverse rendering model (LIRM) based on Transformer, capable of simultaneously reconstructing high-quality geometric shapes, material properties, and view-dependent radiance fields from a small number (3-8) of perspective images in less than one second of inference time.
LiSu: A Dataset and Method for LiDAR Surface Normal Estimation
Dušan Malić (Graz University of Technology), Horst Possegger (Graz University of Technology)
Domain AdaptationAutonomous DrivingTransformerSupervised Fine-TuningPoint Cloud
🎯 What it does: A LiSu synthetic LiDAR dataset is proposed along with a single-step surface normal estimation method based on graph total variation regularization, which takes into account spatial and temporal consistency.
LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors
Han Zhou (McMaster University), Jun Chen (McMaster University)
GenerationData SynthesisDepth EstimationGaussian SplattingImage
🎯 What it does: A no-reference, illumination-insensitive 3D view synthesis method called LITA-GS is proposed, which can directly generate high-quality, normally exposed view images from multi-view sRGB images in low-light or overexposed environments.
LiveCC: Learning Video LLM with Streaming Speech Transcription at Scale
Joya Chen (National University of Singapore), Mike Zheng Shou (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextBenchmarkAudio
🎯 What it does: Using a large amount of automatic speech recognition (ASR) transcription data, we conduct large-scale pre-training of a video large language model (Video LLM) and subsequently perform fine-grained, real-time speech-video alignment learning.
LiVOS: Light Video Object Segmentation with Gated Linear Matching
Qin Liu (University of North Carolina Chapel Hill), Lijuan Wang (Microsoft)
SegmentationVision Language ModelVideo
🎯 What it does: A lightweight video object segmentation network LiVOS is proposed, using linear matching instead of traditional softmax matching, achieving a constant size memory state.
LLaVA-Critic: Learning to Evaluate Multimodal Models
Tianyi Xiong (University of Maryland), Chunyuan Li (ByteDance)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: This paper presents LLaVA-Critic, an open-source large-scale multimodal model (LMM) that can serve as a general evaluator, scoring the responses of multimodal models, ranking them, and generating reward signals for reinforcement learning.
LLaVA-ST: A Multimodal Large Language Model for Fine-Grained Spatial-Temporal Understanding
Hongyu Li (Beihang University), Si Liu (Beihang University)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: LLaVA-ST is proposed, a multimodal large language model capable of simultaneously handling spatiotemporal fine-grained localization and description of videos.
LLAVIDAL: A Large LAnguage VIsion Model for Daily Activities of Living
Dominick Reilly (University of North Carolina Charlotte), Srijan Das (University of North Carolina Charlotte)
RecognitionObject DetectionObject TrackingTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes a multimodal large language visual model for daily living activities (ADL) called LLAVIDAL, and constructs the ADL-X dataset through a semi-automated process, utilizing three modalities: RGB, skeleton, and human-object interaction to achieve finer-grained action understanding.
LLM-driven Multimodal and Multi-Identity Listening Head Generation
Peiwen Lai (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
GenerationData SynthesisTransformerLarge Language ModelTextMultimodalityAudio
🎯 What it does: Generate natural listener head movements in dialogue scenarios, integrating speech content, acoustic features, and emotional information.
LLMDet: Learning Strong Open-Vocabulary Object Detectors under the Supervision of Large Language Models
Shenghao Fu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes an open vocabulary object detector LLMDet trained under the supervision of large language models, enhancing detection performance through image-level detailed descriptions and region-level phrase generation.
LMO: Linear Mamba Operator for MRI Reconstruction
Wei Li (Zhejiang University of Technology), Jianwei Zheng (Zhejiang University of Technology)
RestorationOptimizationImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A linear Mamba operator (LMO) is proposed for MRI reconstruction, achieving the mapping from undersampled k-space to high-quality images.
Locality-Aware Zero-Shot Human-Object Interaction Detection
Sanghyun Kim (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
RecognitionObject DetectionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A framework called LAIn is proposed for zero-shot human-object interaction (HOI) detection, which enhances the recognition capability of unseen interaction categories by improving the locality and interaction awareness of CLIP.
Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation
Byung Hyun Lee (Seoul National University), Se Young Chun (Seoul National University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper studies a training-free local concept elimination method called GLoCE, which can accurately remove target concepts appearing in text prompts while keeping the rest of the image intact.
Localizing Events in Videos with Multimodal Queries
Gengyuan Zhang (Ludwig Maximilian University of Munich), Jindong Gu (University of Oxford)
Object DetectionRetrievalTransformerLarge Language ModelVision Language ModelDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes a video event localization task aimed at multimodal queries and creates the ICQ benchmark and ICQ-Highlight evaluation set, exploring how to transfer existing natural language query models to multimodal queries.
Locally Orderless Images for Optimization in Differentiable Rendering
Ishit Mehta (University of California San Diego), Ravi Ramamoorthi (University of California San Diego)
OptimizationImage
🎯 What it does: A reverse rendering optimization method using Local Orderless Images (LOI) for multi-scale histogram matching in differentiable rendering is proposed.
LOCORE: Image Re-ranking with Long-Context Sequence Modeling
Zilin Xiao (Rice University), Vicente Ordonez (Czech Technical University in Prague)
RetrievalTransformerImage
🎯 What it does: A list-based local feature re-ranking model LOCORE based on long-context Transformer is proposed.
LOD-GS: Achieving Levels of Detail using Scalable Gaussian Soup
Jianxiong Shen (Tencent AI Lab), Xiaohang Zhan (Tencent AI Lab)
GenerationData SynthesisOptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: In this study, the authors propose a scalable Gaussian Soup representation method that organizes 3D Gaussian voxels using discrete triangle primitives, thereby achieving efficient level of detail (LOD) management.
LOGICZSL: Exploring Logic-induced Representation for Compositional Zero-shot Learning
Peng Wu (Shandong University), Wenguan Wang (University of Macau)
ClassificationRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: The LOGICZSL framework is proposed, which incorporates relational knowledge generated by large language models into the training of the CZSL model through logical rules.
Logits DeConfusion with CLIP for Few-Shot Learning
Shuo Li (Xidian University), Wenping Ma (Xidian University)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Proposes the Logits DeConfusion method, which combines Multi-layer Adapter Fusion (MAF) and Inter-Class Deconfusion (ICD) modules to address the inter-class confusion problem in CLIP for few-shot learning.
LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds
Zihui Zhang (Hong Kong Polytechnic University), Bo Yang (Hong Kong Polytechnic University)
SegmentationKnowledge DistillationPoint Cloud
🎯 What it does: The LogoSP method is proposed to achieve unsupervised 3D point cloud semantic segmentation, training directly on unannotated data.
LoKi: Low-dimensional KAN for Efficient Fine-tuning Image Models
Xuan Cai (Shanghai Jiao Tong University), Hua Yang (Shanghai Jiao Tong University)
ClassificationRecognitionTransformerSupervised Fine-TuningImageVideo
🎯 What it does: A low-dimensional KAN adapter named LoKi has been designed and validated for parameter-efficient fine-tuning in pre-trained visual models.
Long Video Diffusion Generation with Segmented Cross-Attention and Content-Rich Video Data Curation
Xin Yan (01.AI), Huan Yang (01.AI)
GenerationTransformerDiffusion modelVideoText
🎯 What it does: Proposes the Presto model, capable of generating 15-second long videos while maintaining long-range coherence and content richness;
LongDiff: Training-Free Long Video Generation in One Go
Zhuoling Li (Lancaster University), Jun Liu (Lancaster University)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: A training-agnostic LongDiff method is proposed, which utilizes position mapping and information frame selection to extend short video generation to long videos.
LongVALE: Vision-Audio-Language-Event Benchmark Towards Time-Aware Omni-Modal Perception of Long Videos
Tiantian Geng (Southern University of Science and Technology), Feng Zheng (Southern University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper presents LongVALE, a comprehensive multimodal (visual, audio, speech) video benchmark consisting of 8.4K long videos, 105K fine-grained temporal boundaries, and multimodal associative descriptions. Based on this benchmark, LongVALE-LLM was trained to achieve cross-modal reasoning and fine-grained temporal understanding of long videos.
LookCloser: Frequency-aware Radiance Field for Tiny-Detail Scene
Xiaoyu Zhang (SenseTime Research), Hujun Bao (Zhejiang University)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: This paper proposes a frequency-aware neural radiance field (FA-NeRF) that can simultaneously reconstruct the overall structure and details of a scene within a single model.
LookingGlass: Generative Anamorphoses via Laplacian Pyramid Warping
Pascal Chang, Vinicius Azevedo
GenerationData SynthesisDiffusion modelFlow-based ModelRectified FlowAuto EncoderImage
🎯 What it does: A generative framework based on implicit flow models and Laplacian pyramid warping (LookingGlass) is proposed, capable of generating blurry visual illusions (anamorphoses) with multi-view parallax in one go, while maintaining high-quality images that can be recognized even in an undistorted state.
LoRA Recycle: Unlocking Tuning-Free Few-Shot Adaptability in Visual Foundation Models by Recycling Pre-Tuned LoRAs
Zixuan Hu (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
Domain AdaptationMeta LearningImageBenchmark
🎯 What it does: The LoRA Recycle framework achieves few-shot adaptation of visual foundation models without fine-tuning by utilizing pre-finetuned LoRA and generating synthetic data through data-free reverse generation, constructing meta-LoRA using meta-learning.
LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning
Xuan Liu (Sun Yat-sen University), Xiaobin Chang (Sun Yat-sen University)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningImage
🎯 What it does: In the scenario of example-free continual learning (EFCL), a drift-resistant space (DRS) is proposed, constructed through the subtraction of LoRA weights, combined with augmented triplet loss to suppress feature drift and retain knowledge.
LoRACLR: Contrastive Adaptation for Customization of Diffusion Models
Enis Simsar (ETH Zurich), Pinar Yanardag (Virginia Tech)
GenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: A method called LoRACLR based on contrastive learning is proposed, which can merge multiple LoRA fine-tuned models into a single model without retraining, thus achieving high-fidelity generation of multi-concept images.
LoRASculpt: Sculpting LoRA for Harmonizing General and Specialized Knowledge in Multimodal Large Language Models
Jian Liang (Wuhan University), Mang Ye (Wuhan University)
TransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: This paper studies and proposes the LoRASculpt framework, which reduces forgetting and enhances performance of multimodal large language models in downstream tasks through sparse sculpting and conflict mitigation regularization of LoRA.
Lost in Translation, Found in Context: Sign Language Translation with Contextual Cues
Youngjoon Jang (Visual Geometry Group, University of Oxford), Andrew Zisserman (Visual Geometry Group, University of Oxford)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: A translation framework driven by LLM is constructed by introducing background descriptions, previous sentence translations, and pseudo-word lists as context in continuous sign language translation.
LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty
Christoforos N. Spartalis (University of Amsterdam), Petros Daras (Centre for Research and Technology Hellas)
Computational EfficiencyKnowledge DistillationData-Centric LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: The LoTUS method is proposed, which achieves forgetting of specified training samples by adjusting the probability distribution in the model output space through entropy regulation, avoiding retraining from scratch.
LotusFilter: Fast Diverse Nearest Neighbor Search via a Learned Cutoff Table
Yusuke Matsui (University of Tokyo)
RetrievalComputational EfficiencyText
🎯 What it does: This paper proposes LotusFilter, a post-processing module based on a pre-constructed cutoff table, designed to quickly generate diverse results after nearest neighbor retrieval.
Low-Biased General Annotated Dataset Generation
Dengyang Jiang (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Utilize diffusion models to generate a low-bias general annotation dataset using only category names, and pre-train a visual backbone network on this dataset.
Low-Rank Adaptation in Multilinear Operator Networks for Security-Preserving Incremental Learning
Huu Binh Ta (Qualcomm), Tung Pham (Qualcomm)
Safty and PrivacyImage
🎯 What it does: This study investigates the introduction of low-rank adaptation in Multilinear Operator Networks (MONet) to support incremental learning under encrypted data.