CVPR 2026 Papers — Page 19
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
LAM: Language Articulated Object Modelers
Yipeng Gao (University of Southern California), Laurent Itti (University of Southern California)
GenerationTransformerLarge Language ModelVision Language ModelTextMesh
🎯 What it does: Propose the LAM system, which generates 3D articulated objects with joint motion directly from text prompts through collaboration between large language models (LLMs) and vision-language models (VLMs);
LaMoGen: Language to Motion Generation Through LLM-Guided Symbolic Inference
Junkun Jiang (Hong Kong Baptist University), Jie Chen (Hong Kong Baptist University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a symbolic motion representation called LabanLite based on Labanotation, and built the LaMoGen framework, enabling large language models to generate interpretable and controllable motions through symbolic reasoning from natural language instructions;
LAMP: Language-Assisted Motion Planning for Controllable Video Generation
Muhammed Burak Kizil (Adobe Research), Duygu Ceylan
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningDiffusion modelVideoTextBenchmark
🎯 What it does: LAMP converts natural language descriptions into controllable 3D objects and camera trajectories to drive video generation.
LAMP: Localization Aware Multi-camera People Tracking in Metric 3D World
Nan Yang (Meta Reality Labs Research), Lingni Ma (Meta Reality Labs Research)
Object TrackingPose EstimationConvolutional Neural NetworkTransformerVideoPoint Cloud
🎯 What it does: Proposes the LAMP method, which first projects 2D keypoints from multiple cameras into a 3D ray cloud using known 6DoF camera poses, then employs a spatiotemporal Transformer to regress human SMPL parameters, achieving continuous multi-person pose tracking in a real 3D world coordinate system.
Landscape-Awareness for Geometric View Diffusion Model
Yan-Ting Chen (National Taiwan University), Chun-Yi Lee (National Taiwan University)
Pose EstimationOptimizationConvolutional Neural NetworkDiffusion modelScore-based ModelImage
🎯 What it does: Estimate the camera pose from two views by leveraging the inverse process of diffusion models, and introduce a two-stage optimization based on landscape analysis.
LangField4D: Learning Identity-Adaptive and Spatio-Temporal Continuous 4D Language Fields for Dynamic Scenes
Yichao Xu (Zhejiang University), Yawei Luo (Zhejiang University)
SegmentationVision Language ModelGaussian SplattingVideoMultimodalityBenchmark
🎯 What it does: Proposes LangField4D, constructing a 4D language field to enable open-source vocabulary queries and semantic awareness in dynamic scenes.
LangRef3DGS: Natural Language-Guided 3D Referential Segmentation from Partial Observations via 3D Gaussian Splatting
Xulun Ye (Ningbo University), Kun Zhou (Shenzhen University)
SegmentationVision Language ModelContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: Construct a real-time language-guided 3D segmentation framework based on 3D Gaussian Splatting, which can perform semantic segmentation and discover new categories under partial RGB-D observations.
Language Does Matter for Cross-Domain Few-Shot Visual Feature Enhancement
Fei Zhou (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
ClassificationObject DetectionSegmentationDomain AdaptationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a cross-modal visual feature enhancement framework that leverages image-level and domain-level language descriptions to guide pre-trained visual models in cross-domain few-shot learning adaptation.
Language Models Can Explain Visual Features via Steering
Javier Ferrando (Barcelona Supercomputing Center), Dario Garcia-Gasulla (Barcelona Supercomputing Center)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelAuto EncoderMultimodality
🎯 What it does: This paper proposes using Vision-Language Models (VLM) to perform causal intervention (Steering) on the sparse autoencoder (SAE) features of the visual encoder, enabling the language model to automatically generate visual concepts represented by these features. It further combines this with the traditional Top-k input example method to form a hybrid explanation framework called Steering-informed Top-k.
Language-driven Fine-grained Retrieval
Shijie Wang (University of Queensland), Zi Huang (University of Adelaide)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes a language-driven fine-grained retrieval framework, LaFG, which utilizes large language models (LLM) to generate attribute-level descriptions for each category. These texts are mapped into the visual space via a visual-language model (VLM), constructing a global attribute vocabulary and aggregating each category into a language prototype. The prototype is then used as a supervisory signal to train the retrieval network, achieving comparable modeling of fine-grained details.
Language-Free Generative Editing from One Visual Example
Omar Elezabi (Computer Vision Lab, CAIDAS & IFI, University of Würzburg), Radu Timofte (Computer Vision Lab, CAIDAS & IFI, University of Würzburg)
RestorationGenerationDiffusion modelScore-based ModelImage
🎯 What it does: Propose a training-free Visual Diffusion Conditioning (VDC) framework that directly learns visual conditions using a single or few visual examples, achieving high-quality image editing through conditional guidance and inversion correction.
Language-Grounded Decoupled Action Representation for Robotic Manipulation
Wuding Weng (Tongji University), Heng Tao Shen (Tongji University)
Robotic IntelligenceTransformerVision-Language-Action ModelContrastive LearningMultimodalitySequential
🎯 What it does: Proposes Language-Grounded Decoupled Action Representation (LaDA), which decomposes 7-DoF control into three motion primitives (translation, rotation, gripper) that can be linguistically described, and uses soft-label contrastive learning and adaptive weighting mechanisms to achieve cross-task action semantic alignment.
Language-guided Frequency Modulation for Large Vision-Language Models
Shuyi Ouyang (Zhejiang University), Xinchao Wang (National University of Singapore)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposes a language-guided frequency-domain modulation (LFM) method, which dynamically adjusts visual features based on textual context through frequency-domain filters without adding extra parameters, achieving hierarchical alignment between vision and language.
Language-Guided One-Step Diffusion Model for Nighttime Flare Removal
Aoxiang Ning (Chongqing University of Technology), Yirui Wu (Hohai University)
RestorationObject DetectionData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Proposed a language-guided single-step diffusion model for nighttime flare removal.
LAOF: Robust Latent Action Learning with Optical Flow Constraints
Xizhou Bu (Fudan University), Wei Li (Northwestern Polytechnical University)
Representation LearningTransformerReinforcement LearningOptical FlowVideo
🎯 What it does: Propose a pseudo-supervised framework called LAOF that leverages optical flow constraints to learn robust latent action representations from unlabeled videos.
Large-scale Codec Avatars: The Unreasonable Effectiveness of Large-scale Avatar Pretraining
Junxuan Li (Meta), Shunsuke Saito (Meta)
GenerationData SynthesisTransformerGaussian SplattingVideo
🎯 What it does: Designed a two-stage (pre-training + post-training) large-scale encoder-decoder framework for one-time generation of high-fidelity, fully controllable 3D full-body human avatars, supporting advanced features such as loose clothing and relighting;
Large-scale Robust Enhanced Ensemble Clustering via Outlier Decoupling
Jiaxuan Xu (Sichuan University), Liang Du (Shanxi University)
Anomaly DetectionComputational EfficiencyGraphBenchmark
🎯 What it does: Propose a robust anchor-based ensemble clustering method called RANGE. First, map the base clustering results to a bipartite graph, enhance the reliability of the bipartite graph using high-order fuzzy enhancement, then decompose the similarity matrix in the anchor space to obtain clean clustering structures and residual anomalous structures. Finally, perform k-means on the clean part to achieve consensus clustering; the residual part can be directly used for anomaly detection.
LaRP: Efficient Multi-View Inpainting with Latent Reprojection Priors
Gaoyang Zhang (Zhejiang University), Xinguo Liu
RestorationConvolutional Neural NetworkDiffusion modelNeural Radiance FieldImagePoint Cloud
🎯 What it does: A framework called LaRP is proposed for multi-view image restoration tasks, ensuring that the restored results maintain 3D consistency across all views.
LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency
Weilong Yan (National University of Singapore), Jingyu Hu (Chinese University of Hong Kong)
RestorationGenerationDiffusion modelPoint CloudMeshBenchmark
🎯 What it does: Propose LaS-Comp, a zero-shot, class-agnostic 3D shape completion framework that leverages geometric priors from pre-trained 3D foundation models to complete shapes from partial observations to full shapes.
LASAR: Towards Spatio-temporal Reasoning with Latent Cognitive Map
Jinzhou Tang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
Autonomous DrivingExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes LASAR, a dual-memory LLM agent that integrates action navigation with spatial reasoning, and constructs an interpretable and transferable spatial cognitive model through ST-CRL contrastive learning on cognitive maps.
LASER: Layer-wise Scale Alignment for Training-Free Streaming 4D Reconstruction
Tianye Ding (Northeastern University), Huaizu Jiang (Northeastern University)
Depth EstimationSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: Convert offline 4D reconstruction models (e.g., VGGT, π³) into an untrained streaming system through sliding window and hierarchical scale alignment, achieving online continuous 3D reconstruction.
LaSM: Layer-wise Scaling Mechanism for Defending Pop-up Attack on GUI Agents
Zihe Yan (Shanghai Jiao Tong University), Gongshen Liu (Shanghai Jiao Tong University)
Adversarial AttackTransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose a post-training hierarchical scaling mechanism called LaSM to enhance the robustness of GUI agents against pop-up injection attacks.
LATA: Laplacian-Assisted Transductive Adaptation for Conformal Uncertainty in Medical VLMs
Behzad Bozorgtabar (Aarhus University), Zongyuan Ge (Monash University)
Domain AdaptationVision Language ModelImageBiomedical Data
🎯 What it does: Proposes LATA, a transductive refinement method that does not require labels or training, for achieving segmented conformal prediction on medical vision-language models (VLMs), improving the efficiency and fairness of confidence sets.
Latent Chain-of-Thought World Modeling for End-to-End Autonomous Driving
Shuhan Tan (University of Texas at Austin), Boris Ivanovic (University of Texas at Austin)
Autonomous DrivingReinforcement LearningVision-Language-Action ModelWorld ModelImageVideoMultimodalityChain-of-Thought
🎯 What it does: Propose LCDrive, a framework that employs latent chain-of-thought (Latent CoT) in end-to-end driving VLA models, directly alternating between generating action suggestions and world model predictions in the latent space, unifying reasoning and decision-making.
Latent Diffusion Inversion Requires Understanding the Latent Space
Mingxing Rao (Vanderbilt University), Daniel Moyer (Vanderbilt University)
Explainability and InterpretabilityAdversarial AttackDiffusion modelAuto EncoderImage
🎯 What it does: Investigated the memorization properties of latent diffusion models (LDMs), discovering uneven memorization in the latent space, and proposed a geometric method based on the decoder pullback metric to measure the contribution of each latent dimension to memorization, thereby enhancing membership inference attack performance by filtering low-contribution dimensions.
Latent Implicit Visual Reasoning
Kelvin Li (University Of California Berkeley), Roei Herzig (University Of California Berkeley)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes the Latent Implicit Visual Reasoning (LIVR) mechanism, which enables models to automatically learn and utilize visual abstractions without explicit intermediate visual supervision by introducing implicit visual attention tokens into large multimodal models (LMMs);
LATTICE: Democratize High-Fidelity 3D Generation at Scale
Zeqiang Lai (CUHK), Xiangyu Yue (CUHK)
GenerationTransformerRectified FlowAuto EncoderImageMeshBenchmark
🎯 What it does: Propose the LATTICE framework, using VoxSet's semi-structured latent representation to generate high-fidelity, editable 3D assets from a single image;
LaVR: Scene Latent Conditioned Generative Video Trajectory Re-Rendering using Large 4D Reconstruction Models
Mingyang Xie (Meta), Lei Luo (Meta)
GenerationData SynthesisDiffusion modelFlow-based ModelAuto EncoderVideo
🎯 What it does: Using the potential space of a pre-trained large 4D reconstruction model (CUT3R) as a soft geometric condition, combined with a video diffusion model, to generate dynamic scene re-rendering for a given new camera trajectory.
Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection
Yawen Yang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
Anomaly DetectionTransformerImage
🎯 What it does: Proposes a synthetic image detection method based on the consistency differences of mid-level features in visual Transformers, called Layer Transition Discrepancy (LTD).
Layer-wise Instance Binding for Regional and Occlusion Control in Text-to-Image Diffusion Transformers
Ruidong Chen (Tianjin University), Anan Liu (Tianjin University)
GenerationTransformerDiffusion modelImageTextMultimodality
🎯 What it does: Propose a zero-shot LayerBind controller that can achieve regional instruction layout and occlusion control for Diffusion Transformer (DiT) models, supporting editable generation
Layered 4D-Rotor Gaussian Splatting: A Compressed Representation for Long Dynamic Scenes
Hanjie Xu (Peking University), He Wang (Galbot)
CompressionComputational EfficiencyNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: Propose Layered 4D-Rotor Gaussian Splatting (L4DRotorGS), a compressed representation for long dynamic scenes that enables real-time rendering while significantly reducing storage and memory requirements.
LayoutAD: Exploring Semantic-Geometric Misalignment Reasoning for Scene Layout Anomaly Detection
Zhichao Zeng (Xidian University), Xiaotian Qiao (Xidian University)
Anomaly DetectionGraph Neural NetworkTransformerImageBenchmark
🎯 What it does: Proposes an unsupervised scene layout anomaly detection framework named LayoutAD, capable of detecting anomalies in object attributes and relationships within images.
LazyVAR: Accelerating Visual Autoregressive Models via Scale-wise Token Pruning and Parallel Group Decoding
Rongge Mao (USTC), S Kevin Zhou (USTC)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: A training-agnostic acceleration method called LazyVAR is proposed for visual autoregressive (VAR) models, which utilizes the similarity of latent features aggregated at the scale level for token pruning and achieves efficient inference through parallel group decoding.
LDP-Slicing: Local Differential Privacy for Images via Randomized Bit-Plane Slicing
Yuanming Cao (McMaster University), Wenbo He (McMaster University)
Safty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: Propose the LDP-Slicing framework, achieving local differential privacy at the pixel level by combining frequency domain low-frequency removal, bit-plane randomization, and budget optimization, enabling privacy-preserving image processing that is friendly to trainable models.
LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving
Long Nguyen (University of Tübingen), Kashyap Chitta (NVIDIA Research)
Autonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: Proposed the LEAD expert and dataset, aligning through visibility, uncertainty, and intention, improving target point conditioning, training TransFuser v6, achieving optimal closed-loop driving performance on CARLA benchmarks.
LEADER: Learning Reliable Local-to-Global Correspondences for LiDAR Relocalization
Jianshi Wu (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Cheng Wang (Fujian Key Laboratory of Urban Intelligent Sensing and Computing)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a LiDAR Absolute Pose Re-localization framework named LEADER, which directly outputs global 6-DoF poses by performing scene coordinate regression on single-frame LiDAR point clouds.
LeapAlign: Post-training Flow Matching Models at Any Generation Step by Building Two-Step Trajectories
Zhanhao Liang (The Australian National University), Liang Zheng (ByteDance Seed)
GenerationFlow-based ModelRectified FlowImageText
🎯 What it does: Propose LeapAlign post-training method for flow-matching models, enabling reward gradients to be directly backpropagated to any generation step, thereby enhancing image generation quality and text consistency.
Learnability-Driven Submodular Optimization for Active Roadside 3D Detection
Ruiyu Mao (University of Texas at Dallas), Yunhui Guo (University of Texas at Dallas)
Object DetectionAutonomous DrivingOptimizationImage
🎯 What it does: Propose an active learning framework LH3D based on learnability for monocular roadside 3D detection, selecting images that are both easy to annotate and informative under a limited annotation budget.
Learnability-Guided Diffusion for Dataset Distillation
Jeffrey A. Chan-Santiago (University of Central Florida), Mubarak Shah (University of Central Florida)
Data SynthesisKnowledge DistillationDiffusion modelImage
🎯 What it does: Propose an incremental dataset distillation framework based on learnability, using diffusion models to generate samples that align with the learning frontier of the current model, thereby constructing a synthetic dataset without redundancy.
Learnable Motion-Focused Tokenization for Effective and Efficient Video Unsupervised Domain Adaptation
Tzu Ling Liu (University of Saskatchewan), Mrigank Rochan (University of Saskatchewan)
Domain AdaptationTransformerReinforcement LearningVision Language ModelVideo
🎯 What it does: In the video unsupervised domain adaptation task, this paper proposes Learnable Motion-Focused Tokenization (LMFT), which learns a threshold to prune low-motion background tokens and retains only motion-rich action-related tokens, thereby enhancing the model's generalization to the target domain and significantly reducing computational costs.
Learned Image Compression via Sparse Attention and Adaptive Frequency
Huidong Ma (Nankai University), Wentong Cai (Nanyang Technological University)
CompressionTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposed a learning-based image compression framework named SAAF, integrating a spatial-frequency dual-path transform network, sparse attention module, content-adaptive frequency blocks, and denoising regularization during training.
Learning 3D Reconstruction with Priors in Test Time
Lei Zhou (Stony Brook University), Dimitris Samaras (Stony Brook University)
Pose EstimationDepth EstimationOptimizationTransformerGaussian SplattingPoint CloudBenchmark
🎯 What it does: This paper proposes a test-time constrained optimization framework that leverages camera pose, intrinsic parameters, and depth priors through loss constraints rather than network inputs, enhancing the performance of multi-view Transformers in 3D reconstruction and camera pose estimation tasks.
Learning 3D Representations for Spatial Intelligence from Unposed Multi-View Images
Bo Zhou (Nanjing University of Science and Technology), Wenguan Wang (Zhejiang University)
SegmentationPose EstimationDepth EstimationKnowledge DistillationRepresentation LearningTransformerVision Language ModelGaussian SplattingImageBenchmark
🎯 What it does: UniSplat proposes an end-to-end forward 3D representation learning framework that can directly learn unified geometric, appearance, and semantic representations from uncalibrated multi-view images.
Learning 3D Shape Fidelity Metric from Real-world Distortions
Xuelu Feng, Chunming Qiao
Representation LearningGraph Neural NetworkSupervised Fine-TuningPoint CloudMeshBenchmark
🎯 What it does: Designed and implemented a learning-based 3D shape fidelity metric called LoCaSE, which is based on a local connectivity attention mechanism, and constructed a real-world reconstruction/generative distortion annotated dataset called RSF, using human evaluations for model training and assessment.
Learning a Unified Latent Action Space from Videos with Action-centric Cycle Consistency
Guangyan Chen (Beijing Institute of Technology), Yufeng Yue (LimX)
Robotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelAuto EncoderContrastive LearningVideo
🎯 What it does: Proposes a unified latent action space based on action-center cyclic consistency for learning robot policies from unannotated videos.
Learning Anchor in Dual Orthogonal Space for Fast Multi-view Clustering
Yalan Qin (Shanghai University), Hanzhou Wu (Shanghai University)
OptimizationComputational EfficiencyRepresentation LearningMultimodality
🎯 What it does: Proposes a multi-view fast clustering method (DOSFMVC) that learns anchor points in a dual orthogonal space, achieving high-quality clustering through joint learning of anchors and clustering indicators.
Learning and Aligning Click-Aware Shape Prior for Interactive Amodal Instance Segmentation
Junjie Chen (Jiangxi University of Finance and Economics), Yifan Zuo (Jiangxi University of Finance and Economics)
SegmentationAutonomous DrivingTransformerContrastive LearningImage
🎯 What it does: Proposes an interactive fully occluded instance segmentation framework called ClickPriorNet, which utilizes click information to retrieve and align shape priors to improve the segmentation quality of occluded regions.
Learning by Analogy: A Causal Framework for Compositional Generalization
Lingjing Kong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
GenerationTransformerSupervised Fine-TuningDiffusion modelMultimodality
🎯 What it does: This paper proposes a hierarchical latent model based on causal modularization and minimum change to explain and achieve compositional generalization in image generation.
Learning Compact 3D Representations from Feed-Forward Novel View Synthesis
Honggyu An (KAIST AI), Seungryong Kim (KAIST AI)
GenerationData SynthesisTransformerGaussian SplattingImageBenchmark
🎯 What it does: This paper proposes a Transformer framework called C3G based on learnable queries, which can learn compact 3D Gaussian representations from sparse multi-view images without poses, and achieves view-invariant feature enhancement through C3GF for high-quality view synthesis and 3D scene understanding.
Learning complete and explainable visual representations from itemized text supervision
Yiwei Lyu (University of Michigan), Todd C. Hollon (University of Michigan)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: Propose the ItemizedCLIP framework by leveraging itemized text supervision, using cross-attention modules to learn complete and interpretable visual representations.
Learning Convex Decomposition via Feature Fields
Yuezhi Yang (NVIDIA), Nicholas Sharp (NVIDIA)
OptimizationRepresentation LearningTransformerContrastive LearningMesh
🎯 What it does: The study proposes an end-to-end model for convex decomposition of 3D shapes by learning feature fields.
Learning Coordinate-based Convolutional Kernels for Continuous SE(3) Equivariant and Efficient Point Cloud Analysis
Jaein Kim (Seoul National University), Byoung-Tak Zhang (Seoul National University)
ClassificationSegmentationPose EstimationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Propose a continuous SE(3) equivariant convolutional network, ECKConv, which parameterizes the convolution kernel in the double coset space to achieve strict equivariance under rigid transformations, and realizes efficient memory and computational scalability.
Learning Cross-View Object Correspondence via Cycle-Consistent Mask Prediction
Shannan Yan (Tsinghua University), Fengyun Rao (WeChat Vision, Tencent Inc)
Object DetectionSegmentationRepresentation LearningTransformerImageVideo
🎯 What it does: Propose a cross-view object correspondence framework based on conditional binary segmentation, achieving unsupervised self-supervised learning through cyclic consistency mask prediction.
Learning Diffeomorphism for Medical Image Registration with Time-Embedded Architectures Using Semigroup Regularization
Mohammadjavad Matinkia (University of Alberta), Nilanjan Ray (University of Alberta)
Convolutional Neural NetworkTransformerDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed TomographyOrdinary Differential Equation
🎯 What it does: Propose a continuous-time diffeomorphic registration framework SGDIR that can learn reversible, topology-preserving deformation flows using only semigroup regularization, eliminating the integration and multiple regularization requirements of traditional methods.
Learning Differentiable Hierarchies in 3D Gaussian Splatting
Youqi Pan (Peking University), Hongbin Zha (Peking University)
CompressionRepresentation LearningGaussian SplattingImage
🎯 What it does: Learning a differentiable hierarchical representation in 3D Gaussian Splatting enables the model to achieve Level of Detail (LoD) rendering and model compression with arbitrary numbers of Gaussians.
Learning Effective Sign Features without Text for Gloss-free Sign Language Translation
Shiwei Gan (Nanjing University), Hongkai Wen (University of Warwick)
RecognitionKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningVideo
🎯 What it does: This paper proposes a fully text-free self-supervised pre-training framework called SignDINO, specifically designed for sign language translation tasks. It learns a visual encoder capable of capturing fine-grained actions and reasoning semantics from global views through self-distillation on both global views and local (facial and hand) views of sign language videos.
Learning Eigenstructures of Unstructured Data Manifolds
Roy Velich (Technion Israel Institute Of Technology), Ron Kimmel
Representation LearningAuto EncoderImagePoint Cloud
🎯 What it does: Propose an unsupervised framework that directly learns spectral bases from unstructured point clouds or high-dimensional data, eliminating the need for explicit operator construction and eigen-decomposition steps.
Learning Explicit Continuous Motion Representation for Dynamic Gaussian Splatting from Monocular Videos
Xuankai Zhang (Sun Yat-sen University), Qing Zhang (Sun Yat-sen University)
GenerationDiffusion modelGaussian SplattingVideo
🎯 What it does: Proposes a method that utilizes adaptive SE(3) B-spline motion bases to explicitly model continuous pose deformations of dynamic Gaussian clouds, achieving high-quality dynamic Gaussian splatting from monocular video;
Learning Forgery-Aware Lip Representations Without Forgery Priors
Bofan Chen (Shanghai Jiao Tong University), Shi-Lin Wang (Shanghai Jiao Tong University)
ClassificationAnomaly DetectionRepresentation LearningContrastive LearningVideo
🎯 What it does: Propose a visual speaker authentication (VSA) detector without prior knowledge of forgeries, trained solely on real samples, capable of identifying forgeries and performing authentication under various talking face forgery (TFG) attacks.
Learning from Itself: Mining Internal Knowledge from Vision Language Models for Continual Learning
Yizheng Gong (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Xi'an Jiaotong-Liverpool University)
Knowledge DistillationTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes the Learning from Itself (LfI) method, enabling CLIP to accomplish continual learning tasks through self-generated pseudo descriptions and adaptive mutual distillation.
Learning from Noisy Supervision: A Denoising-Debiasing Framework for Weakly Supervised Video Anomaly Detection
Yaxin Zhao (Nankai University), Xiaojie Yuan (Shanghai Jiao Tong University)
Anomaly DetectionVision Language ModelContrastive LearningVideo
🎯 What it does: In weakly supervised video anomaly detection, the D²MIL framework is proposed, which combines dynamic threshold denoising and vision-language model debiasing to suppress noise samples and retain hard anomaly samples during the multi-instance learning process.
Learning from Oblivion: Predicting Knowledge-Overflowed Weights via Retrodiction of Forgetting
Jinhyeok Jang (ETRI), Jung Uk Kim (Kyung Hee University)
ClassificationSegmentationDomain AdaptationMeta LearningImage
🎯 What it does: This paper proposes a method to predict richer pre-trained weights (KNOW prediction) by utilizing structured progressive forgetting and its inverse process, and designs a lightweight meta-learning hypernetwork called KNOWN.
Learning from Semantic Dictionaries: Discriminative Codebook Contrastive Learning for Unified Visual Representation and Generation
Imanol G. Estepa (Universitat de Barcelona), Petia Radeva (Universitat de Barcelona)
GenerationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose the LEASEn framework in unsupervised pre-training, leveraging paired codebooks from generative and discriminative models to jointly learn visual representations and image generation;
Learning from Synthetic Data via Provenance-Based Input Gradient Guidance
Koshiro Nagano (Keio University), Hideo Saito (Keio University)
ClassificationData SynthesisExplainability and InterpretabilityDiffusion modelImageVideo
🎯 What it does: Utilize the automatically obtained provenance information during the synthetic data generation process as auxiliary supervision, employing input gradient guidance (provenance loss) to suppress the model's dependence on non-target regions, thereby enhancing discriminative ability for the target region.
Learning Generalizable 3D Medical Image Representations from Mask-Guided Self-Supervision
Yunhe Gao (Stanford University), Curtis Langlotz (Stanford University)
SegmentationRepresentation LearningConvolutional Neural NetworkContrastive LearningMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: Developed and validated the MASS (Mask-guided Self-Supervised learning) framework, which performs self-supervised pretraining on 3D medical images using automatically generated class-agnostic masks, thereby learning transferable, semantically rich, and spatially precise features.
Learning Hierarchical Hyperbolic Mixture Model for Part-aware 3D Generation
Qitong Yang (Xidian University), Ajmal Mian (University of Western Australia)
GenerationDiffusion modelGaussian SplattingPoint CloudMeshStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a 3D shape generation framework based on a hyperspherical hierarchical hybrid model, achieving part-aware and hierarchical high-quality 3D generation.
Learning Latent Concepts for Detecting Out-of-Distribution Objects
Ting Peng (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
Object DetectionDomain AdaptationAnomaly DetectionTransformerContrastive LearningImage
🎯 What it does: Propose the UNO-Adapter framework, which achieves detection of OOD objects without modifying the original detector, through unsupervised concept discovery and neural concept binding.
Learning Latent Proxies for Controllable Single-Image Relighting
Haoze Zheng (Hong Kong University of Science and Technology), Harry Yang (Hong Kong University of Science and Technology)
Image TranslationGenerationDiffusion modelAuto EncoderImage
🎯 What it does: The paper proposes the LightCtrl framework, which achieves controllable relighting from a single image by incorporating lightweight physical proxy encoding and light-aware mask fusion into diffusion models.
Learning Latent Transmission and Glare Maps for Lens Veiling Glare Removal
Xiaolong Qian (Zhejiang University), Kaiwei Wang (Zhejiang University)
RestorationDomain AdaptationDiffusion modelImagePhysics Related
🎯 What it does: This paper proposes a physics-informed generative model, VeilGen, and a reversible restoration network, DeVeiler, for simultaneously correcting diffraction distortion and glare scattering in simplified optical systems, restoring clear images.
Learning Like Humans: Analogical Concept Learning for Generalized Category Discovery
Jizhou Han (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
ClassificationRecognitionRepresentation LearningTransformerContrastive LearningMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed a text concept generation module called ATCG based on analogical reasoning, which helps the model build bridges between known and unknown categories, thereby enhancing the performance of generic category discovery (GCD).
Learning Long-term Motion Embeddings for Efficient Kinematics Generation
Nick Stracke (LMU), Björn Ommer (LMU)
GenerationTransformerVision-Language-Action ModelFlow-based ModelAuto EncoderVideoText
🎯 What it does: Construct a two-stage framework: first, use a VAE to compress sparse tracker trajectories into a 64× temporal compressed dense motion space, then train a conditional flow matching model in this space to achieve long-term motion generation conditioned on text or spatial 'stamps'.
Learning Multi-View Spatial Reasoning from Cross-View Relations
Suchae Jeong (KAIST), Kimin Lee (KAIST)
Data SynthesisRobotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageMultimodalityTime SeriesBenchmark
🎯 What it does: Proposed and constructed the XVR dataset, aiming to enhance multi-perspective spatial reasoning ability through cross-perspective relationship supervision.
Learning Mutual View Information Graph for Adaptive Adversarial Collaborative Perception
Yihang Tao (Hong Kong JC STEM Lab of Smart City), Yuguang Fang (Hong Kong JC STEM Lab of Smart City)
Autonomous DrivingAdversarial AttackRecurrent Neural NetworkGraph Neural NetworkPoint CloudGraphSequential
🎯 What it does: This study proposes an adaptive collaborative perception adversarial attack framework called MVIG, which leverages mutual perspective information graphs (MVIG) to capture vulnerabilities leaked by defensive systems, enabling dynamic attacks on collaborative perception systems.
Learning Personalized Photographic Style from Pairwise User Preferences
Jinwoo Kim (Yonsei University), Seon Joo Kim (Yonsei University)
Image HarmonizationTransformerImage
🎯 What it does: This paper proposes the Personalized Photography Style Learning (PPS) task, which learns and applies personalized color and tone adjustments by collecting users' paired preferences.
Learning Scene Coordinate Reconstruction from Unposed Images via Pose Graph Optimization
Tze Ho Elden Tse (dConstruct Robotics), Angela Yao (National University of Singapore)
Pose EstimationOptimizationSimultaneous Localization and MappingOptical FlowImage
🎯 What it does: Propose a hybrid framework that integrates pose graph optimization (PGO) into ACE-Zero for learning scene coordinate reconstruction from unlabelled images, achieving global consistency through PGO.
Learning Spatial-Temporal Consistency for 3D Semantic Scene Completion
Yujie Xue (Hunan University), Kenli Li (Hunan University)
SegmentationDepth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerImagePoint CloudBenchmark
🎯 What it does: Propose a camera-based 3D semantic scene completion framework named ConSSC, achieving spatiotemporal consistency by leveraging 2D visibility of historical frames and 3D voxel similarity;
Learning Straight Flows: Variational Flow Matching for Efficient Generation
Chenrui Ma (University of California, Irvine), Yanning Shen (University of California, Irvine)
GenerationTransformerFlow-based ModelRectified FlowAuto EncoderImageOrdinary Differential Equation
🎯 What it does: This paper proposes Straight Variational Flow Matching (S-VFM), which introduces variational latent variables and linearity constraints into the Flow Matching framework, enabling generated trajectories to approximate straight lines and achieving generation in one or a few steps;
Learning Surgical Robotic Manipulation with 3D Spatial Priors
Yu Sheng (University Of Science And Technology Of China), Jianmin Ji
Robotic IntelligenceTransformerSupervised Fine-TuningImageVideoPoint Cloud
🎯 What it does: Propose an end-to-end visual-driven surgical robot control strategy—Spatial Surgical Transformer (SST), achieving precise control by learning 3D spatial priors
Learning to Act Robustly with View-Invariant Latent Actions
Youngjoon Jeong (Seoul National University), Taesup Kim (Seoul National University)
Representation LearningRobotic IntelligenceVision-Language-Action ModelContrastive LearningImage
🎯 What it does: Propose a VILA framework that learns view-invariant dynamic representations by performing contrastive learning and future prediction on view-invariant latent actions;
Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration
Weile Chen (Zhejiang University), Siliang Tang (Zhejiang University)
Large Language ModelReinforcement LearningAgentic AIVision-Language-Action ModelTextBenchmark
🎯 What it does: Propose the SCALE framework, which employs self-adversarial Selector, Predictor, and Judger to enable adaptive learning for Web Agents, achieves global exploration through SCALE-Hop, and constructs the SCALE-20k dataset.
Learning to Assist: Physics-Grounded Human-Human Control via Multi-Agent Reinforcement Learning
Yuto Shibata (Carnegie Mellon University), Katerina Fragkiadaki (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningPhysics Related
🎯 What it does: Through a multi-agent reinforcement learning framework, jointly train physics-based control strategies for supporters and recipients, enabling realistic support actions in high-contact, human-robot interaction scenarios;
Learning to Control Physically-simulated 3D Characters via Generating and Mimicking 2D Motions
Jianan Li (Chinese University of Hong Kong), Tien-Tsin Wong (Monash University)
GenerationPose EstimationTransformerReinforcement LearningAuto EncoderVideoPoint Cloud
🎯 What it does: Directly learn a physics simulation character controller using 2D videos to map 2D motion sequences to 3D physically feasible motions.
Learning to Diversify and Focus: A Reinforcement Framework for Open-Vocabulary HOI Detection
Yongchao Xu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RecognitionObject DetectionTransformerReinforcement LearningVision Language ModelImage
🎯 What it does: Propose a Semantic Diversified and Interaction-Focused (SD-IF) framework based on reinforcement learning for open-vocabulary human-object interaction (OV-HOI) detection, addressing the issues of overfitting to seen interactions and insufficient spatial perception of interaction details in traditional single-stage methods.
Learning to Drive is a Free Gift: Large-Scale Label-Free Autonomy Pretraining from Unposed In-The-Wild Videos
Matthew Strong (Applied Intuition), Wei Zhan (Applied Intuition)
Autonomous DrivingRepresentation LearningData-Centric LearningTransformerWorld ModelVideo
🎯 What it does: Leverages unlabeled, unposed driving videos combined with teacher-guided unsupervised learning to construct a unified pseudo 4D geometric, semantic, and motion representation, achieving short-term prediction of future frames.
Learning to Focus and Precise Cropping:A Reinforcement Learning Framework with Information Gaps and Grounding Loss for MLLMs
Xuanpu Zhao (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
Convolutional Neural NetworkLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Proposes a two-stage reinforcement learning framework based on information gap and localization loss, enabling multi-modal large language models to actively focus and precisely crop regions of interest in high-resolution visual question answering.
Learning to Generate Highly Dynamic Videos using Synthetic Motion Data
Wonjoon Jin (POSTECH), Sunghyun Cho (POSTECH)
GenerationData SynthesisDiffusion modelFlow-based ModelAuto EncoderOptical FlowVideo
🎯 What it does: Designed and implemented a two-stage video synthesis framework based on diffusion models—DynaVid—which first generates optical flow (i.e., motion patterns) through a motion generator, and then maps the optical flow to RGB videos via a motion-guided video generator, enabling controllable generation of highly dynamic object motion and extreme camera movements.
Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models
Jiadong Pan (Institute Of Computing Technology Chinese Academy Of Sciences), Haifeng Wang (Baidu Inc)
GenerationTransformerReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: Propose a token-level internal reward mechanism called GvU, which leverages the understanding branch of a unified multimodal model to enhance text-to-image generation quality through self-supervised reinforcement learning.
Learning to Identify Out-of-Distribution Objects for 3D LiDAR Anomaly Segmentation
Simone Mosco (University of Padova), Alberto Pretto (University of Padova)
SegmentationAnomaly DetectionConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: Proposes an efficient method for 3D LiDAR anomaly segmentation by directly modeling the in-class distribution in the feature space;
Learning to Infer Parameterized Representations of Plants from 3D Scans
Samara Ghrer (Inria Center at Universit Grenoble Alpes), Stefanie Wuhrer (Inria Center at Universit Grenoble Alpes)
SegmentationGenerationData SynthesisRepresentation LearningRecurrent Neural NetworkPoint CloudAgriculture Related
🎯 What it does: Infer a parameterized L-String representation of plants from 3D point clouds and use this representation to accomplish tasks such as reconstruction, skeleton extraction, and segmentation.
Learning to Learn Weight Generation via Local Consistency Diffusion
Yunchuan Guan (Huazhong University of Science and Technology), Lei Li (Huazhong University of Science and Technology)
Meta LearningConvolutional Neural NetworkDiffusion modelImageText
🎯 What it does: This paper proposes a weight generation framework called Mc-Di, based on meta-learning and local consistency diffusion, for rapidly generating network weights applicable to multi-task scenarios.
Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models
Shengchao Zhou (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoPoint CloudBenchmark
🎯 What it does: Built a dynamic spatial reasoning (DSR) dataset and evaluation framework, and improved vision-language models (VLM) to enhance reasoning capabilities for 4D dynamic scenes.
Learning to Refuse: Refusal-Aware Reinforcement Fine-Tuning for Hard-Irrelevant Queries in Video Temporal Grounding
Jin-Seop Lee (Sungkyunkwan University), Jee-Hyong Lee (Sungkyunkwan University)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelVideo
🎯 What it does: This paper proposes a post-training method based on reinforcement learning (RA-RFT), which enables video temporal localization models to make rejection decisions and provide explanations when faced with semantically similar but irrelevant queries.
Learning to See and Act: Task-Aware Virtual View Exploration for Robotic Manipulation
Yongjie Bai (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
Robotic IntelligenceTransformerReinforcement LearningMixture of ExpertsVision-Language-Action ModelPoint CloudBenchmark
🎯 What it does: Proposes the Task-Aware Virtual Perspective Exploration (TVVE) framework, which achieves more robust robot manipulation by dynamically selecting the most informative perspectives and using a multi-task visual encoder.
Learning to See Through a Baby's Eyes: Early Visual Diets Enable Robust Visual Intelligence in Humans and Machines
Yusen Cai (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
ClassificationRecognitionDepth EstimationTransformerContrastive LearningImageVideo
🎯 What it does: Proposed self-supervised learning training strategies based on infant visual development stages (CATDiet and CombDiet), evaluated on ten visual tasks.
Learning to See through Illumination Extremes with Event Streaming in Multimodal Large Language Models
Baoheng Zhang (University of Hong Kong), Hayden Kwok-Hay So (University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose Event-MLLM, which integrates event camera data with RGB to enable multimodal reasoning and instruction following under extreme lighting conditions.
Learning to Select Visual Tools from Experience
Zeyi Huang (Microsoft), Yong Jae Lee (University of Illinois Urbana-Champaign)
Reinforcement LearningAgentic AIPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes VisualToolAgent (VisTA), a reinforcement learning framework that enables visual agents to autonomously explore, select, and combine diverse tools without relying on human annotations, assisting frozen vision-language models in visual reasoning tasks.
Learning to Solve PDEs on Neural Shape Representations
Lilian Welschinger (University College London), Niloy J. Mitra (University College London)
Point CloudMeshPhysics Related
🎯 What it does: This paper proposes a mesh-free, geometry-conditioned neural PDE solver capable of directly solving surface partial differential equations within neural surface representations (such as point clouds, implicit fields, Spherical Neural Surfaces, etc.).
Learning to Track Instance from Single Nature Language Description
Yaozong Zheng (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
Object TrackingTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposed a self-supervised visual language tracking method called SVLTrack that uses only natural language descriptions
Learning Transferable Temporal Primitives for Video Reasoning via Synthetic Videos
Songtao Jiang (Zhejiang University), Zuozhu Liu (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: Proposes the SynRL framework, which teaches visual language models (VLM) fundamental temporal primitives (direction, speed, state tracking, etc.) using procedurally generated synthetic videos and precise frame-level annotations, and achieves post-training enhancement through chain-of-thought (CoT) and reinforcement learning (GRPO).
Learning What Helps: Task-Aligned Context Selection for Vision Tasks
Jingyu Guo (Kth Royal Institute Of Technology), Kevin Smith (Kth Royal Institute Of Technology)
ClassificationSegmentationTransformerReinforcement LearningImageBiomedical Data
🎯 What it does: Learn how Vision models (e.g., ViT) select context images most beneficial for the task from a candidate pool, integrating the retrieval process as part of the model.