CVPR 2025 Papers — Page 25
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
Split Adaptation for Pre-trained Vision Transformers
Lixu Wang (Nanyang Technological University), Qi Zhu (Northwestern University)
ClassificationDomain AdaptationSafty and PrivacyTransformerImageOrdinary Differential Equation
🎯 What it does: A Split Adaptation method is proposed, which can fine-tune pre-trained visual Transformers with few samples while maintaining data and model privacy.
SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking
Wenrui Cai (Beihang University), Yunhong Wang (Beihang University)
Object TrackingTransformerMixture of ExpertsVideo
🎯 What it does: A visual tracker SPMTrack based on Mixture of Experts (TMoE) is proposed, which can adaptively handle different relationship modeling and introduce spatiotemporal context;
Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving
Alexey Nekrasov (RWTH Aachen University), Julie Stephany Berrio Perez (The University of Sydney)
SegmentationAnomaly DetectionAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: This paper presents the STU dataset, focusing on LiDAR 3D anomaly segmentation, and provides synchronized camera data with dense labels.
SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization
Junchen Yu (Ningbo Innovation Center, Zhejiang University), Hui-Liang Shen (Zhejiang University)
Domain AdaptationOptimizationKnowledge DistillationTransformerImageMultimodality
🎯 What it does: Proposes SSHNet, a framework that transforms unsupervised cross-modal homography estimation into two directly supervised subproblems and jointly trains through segmentation optimization.
STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks
Tianqing Zhang (Zhejiang University), Qiang Zhang (Dalian University of Technology)
ClassificationRecognitionComputational EfficiencySpiking Neural NetworkSupervised Fine-TuningImageTime Series
🎯 What it does: A framework for spiking neural networks (SNN) based on a spatial-temporal attention aggregator (STAA) and time-step random dropout (TSRD) is proposed to enhance the accuracy and efficiency of SNNs.
Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning
Shouhang Zhu (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
OptimizationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A deep reinforcement learning framework based on expert trajectory regularization has been designed and implemented for automatic camera focusing (AF). The labels of the single-step prediction network have been improved to relative motion labels to enhance multi-step focusing accuracy and reduce the phenomenon of focus hunting.
Stable Flow: Vital Layers for Training-Free Image Editing
Omri Avrahami (Snap Research), Daniel Cohen-Or (Snap Research)
GenerationData SynthesisTransformerFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: A training-free image editing method called Stable Flow is proposed, utilizing the 'key layers' of the flow-based Transformer (DiT) model to achieve non-rigid editing, object addition/replacement, and panoramic editing.
Stable-SCore: A Stable Registration-based Framework for 3D Shape Correspondence
Haolin Liu (Chinese University of Hong Kong Shenzhen), Xiaoguang Han (Chinese University of Hong Kong Shenzhen)
SegmentationOptimizationDiffusion modelPoint CloudMeshBenchmark
🎯 What it does: Proposes the Stable-SCore framework, which utilizes 2D foundational models to generate semantic flows and guides Neural Jacobian Fields for differentiable registration of source meshes, achieving stable 3D shape correspondence.
StableAnimator: High-Quality Identity-Preserving Human Image Animation
Shuyuan Tu (Fudan University), Zuxuan Wu (Fudan University)
GenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: We propose StableAnimator, an end-to-end identity-preserving human image animation diffusion framework that directly generates high-quality, identity-consistent videos based on reference images and pose sequences.
Stacking Brick by Brick: Aligned Feature Isolation for Incremental Face Forgery Detection
Jikang Cheng (Wuhan University), Zhongyuan Wang (Wuhan University)
ClassificationRecognitionAnomaly DetectionKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningImageBenchmark
🎯 What it does: This paper proposes an incremental facial forgery detection method that performs Aligned Feature Isolation in the latent space, utilizing Sparse Uniform Replay (SUR) and Latent Incremental Detector (LID) to achieve knowledge retention and new task learning.
StageDesigner: Artistic Stage Generation for Scenography via Theater Scripts
Zhaoxing Gan (Fudan University), Zuo Hu (Shanghai Theatre Academy)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: Based on the automatic generation of foreground 3D objects and emotionally appropriate 2D backgrounds, stage scene generation is achieved.
Star with Bilinear Mapping
Zelin Peng (Shanghai Jiaotong University), Wei Shen (Shanghai Jiaotong University)
ClassificationObject DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: A Transformer-style visual model SBM based on star-shaped operations and low-rank bilinear mapping is proposed, achieving global context modeling with linear complexity.
STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds
Zikuan Li (Nanjing University of Aeronautics and Astronautics), Jun Wang (Nanjing University of Aeronautics and Astronautics)
OptimizationPoint Cloud
🎯 What it does: The STAR-Edge method is proposed, which constructs a structure-aware neighborhood through local spherical curves for edge detection and optimization of thin-walled structure point clouds.
StarGen: A Spatiotemporal Autoregression Framework with Video Diffusion Model for Scalable and Controllable Scene Generation
Shangjin Zhai (SenseTime Research), Guofeng Zhang (State Key Lab of CAD and CG Zhejiang University)
GenerationPose EstimationDiffusion modelGaussian SplattingVideo
🎯 What it does: The StarGen framework is proposed to generate long-range scenes using a video diffusion model under spatial-temporal autoregressive conditions, achieving precise pose control.
StarVector: Generating Scalable Vector Graphics Code from Images and Text
Juan A. Rodriguez (ServiceNow Research), Marco Pedersoli (ServiceNow Research)
GenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A multimodal large language model called StarVector is proposed, which can directly generate SVG code from images or text, achieving high-quality image vectorization to SVG and text-driven SVG generation.
STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction
Zhimin Liao (Xi'an Jiaotong University), Ziyang Ren (Xi'an Jiaotong University)
Object DetectionSegmentationAutonomous DrivingTransformerOptical FlowPoint Cloud
🎯 What it does: Designed and implemented the STCOcc framework, utilizing explicit occupancy states for spatial and temporal cascading reconstruction of 3D voxel features, completing 3D occupancy prediction and scene flow prediction tasks.
STDD: Spatio-Temporal Dual Diffusion for Video Generation
Shuaizhen Yao (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: This paper proposes and implements a spatiotemporal dual diffusion model (STDD), which achieves video generation and prediction by explicitly propagating and denoising in both spatial and temporal dimensions simultaneously.
StdGEN: Semantic-Decomposed 3D Character Generation from Single Images
Yuze He (Tencent AI Lab), Xiao Han (Tencent AI Lab)
GenerationTransformerDiffusion modelImageMesh
🎯 What it does: Generate separable high-quality 3D character models that support rapid synthesis and subsequent editing from a single image;
Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation
Qinghe Ma (Nanjing University), Yinghuan Shi (Nanjing University)
SegmentationDomain AdaptationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingUltrasound
🎯 What it does: A collaborative training framework called SynFoC is proposed, which integrates a foundational model (MedSAM) with a traditional model (U-Net) to address the issues of domain shift and label scarcity in mixed-domain semi-supervised medical image segmentation (MiDSS).
Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for Large Vision Language Models
Zhaoyi Liu (University of Illinois Urbana-Champaign), Huan Zhang (University of Illinois Urbana-Champaign)
Representation LearningAdversarial AttackLarge Language ModelVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: We propose BADVISION, an attack framework that implants covert backdoors into visual encoders using unsupervised learning, capable of inducing misleading hallucinations in large visual language models.
Steepest Descent Density Control for Compact 3D Gaussian Splatting
Peihao Wang (University of Texas at Austin), Rakesh Ranjan (Meta Reality Labs)
OptimizationGaussian SplattingPoint Cloud
🎯 What it does: A theory-driven dense control method called SteepGS is proposed to efficiently escape saddle points and reduce redundant Gaussian points in 3D Gaussian scattering (3DGS).
Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks
Han Wang (University of Illinois Urbana-Champaign), Huan Zhang (University of Illinois Urbana-Champaign)
Adversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper studies a defense method called ASTRA based on adaptive activation vectors, which utilizes visual attribution to identify attack features and dynamically shifts model activations during inference to resist jailbreak attacks on visual language models.
STEP: Enhancing Video-LLMs' Compositional Reasoning by Spatio-Temporal Graph-guided Self-Training
Haiyi Qiu (Zhejiang University), Tat-Seng Chua (National University of Singapore)
GenerationData SynthesisOptimizationTransformerLarge Language ModelVision Language ModelVideoChain-of-Thought
🎯 What it does: Proposes the STEP framework, which uses self-supervised generation of multi-step reasoning question-answer data from video spatiotemporal scene graphs and performs self-training on Video-LLM.
STEPS: Sequential Probability Tensor Estimation for Text-to-Image Hard Prompt Search
Yuning Qiu (RIKEN Center for Advanced Intelligence Project), Qibin Zhao (RIKEN Center for Advanced Intelligence Project)
GenerationOptimizationAdversarial AttackPrompt EngineeringDiffusion modelImageText
🎯 What it does: A new Sequential Tensor Estimation Framework (STEPS) is proposed to optimize discrete hard prompts in text-to-image diffusion models, making the model generate images that are more similar to the target image.
Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail
Luca Bartolomei, Stefano Mattoccia
Depth EstimationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes Stereo Anywhere, a dual-branch deep learning stereo matching framework that integrates stereo geometric constraints with monocular depth Vision Foundation Model (VFM) priors.
STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion Models
Koushik Srivatsan (Johns Hopkins University), Karthik Nandakumar (MBZUAI)
GenerationData SynthesisAdversarial AttackDiffusion modelImageText
🎯 What it does: This paper proposes a two-stage framework named STEREO for achieving robust erasure of unwanted concepts in pre-trained text-to-image diffusion models.
Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos
Linyi Jin (Google DeepMind), Aleksander Holynski
Object TrackingGenerationData SynthesisDepth EstimationTransformerSimultaneous Localization and MappingOptical FlowVideoPoint Cloud
🎯 What it does: A pipeline for automatically extracting pseudo-metric 4D reconstruction (3D structure + long-term motion trajectory) from internet VR180 stereo videos is proposed, and based on this, the DynaDUSt3R model is trained to predict a complete 3D point cloud and its temporal evolution motion trajectory from two frames of images.
StickMotion: Generating 3D Human Motions by Drawing a Stickman
Tao Wang (Beijing University of Posts and Telecommunications), Lei Jin (Beijing University of Posts and Telecommunications)
GenerationData SynthesisPose EstimationTransformerDiffusion modelImageVideo
🎯 What it does: We designed StickMotion, a multi-conditional human motion generation framework based on diffusion models, which can generate 3D motions that align with intentions based on text descriptions and user-drawn stickmen.
STiL: Semi-supervised Tabular-Image Learning for Comprehensive Task-Relevant Information Exploration in Multimodal Classification
Siyi Du (Imperial College London), Chen Qin (Imperial College London)
ClassificationTransformerContrastive LearningImageMultimodalityTabular
🎯 What it does: Proposes the STiL framework, which uses semi-supervised learning to address the modality information gap in image-table multimodal classification.
STING-BEE: Towards Vision-Language Model for Real-World X-ray Baggage Security Inspection
Divya Velayudhan (Khalifa University of Science and Technology), Naoufel Werghi (University of Bonn)
ClassificationRecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A multi-modal X-ray luggage inspection dataset, STray, has been developed, and based on it, a visual language assistant, STING-BEE, has been trained for scene understanding, threat localization, visual grounding, and VQA.
STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation
Yisi Luo (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
RestorationSegmentationData SynthesisAuto EncoderBiomedical Data
🎯 What it does: This paper proposes a STINR framework based on Implicit Neural Representation (INR) for continuous modeling of spatial transcriptomics data, achieving tasks such as gene imputation, denoising, spatial domain detection, and cell type deconvolution.
Stochastic Human Motion Prediction with Memory of Action Transition and Action Characteristic
Jianwei Tang (Sun Yat-sen University), Jian-Fang Hu (Sun Yat-sen University)
GenerationPose EstimationRecurrent Neural NetworkAuto EncoderVideo
🎯 What it does: This study investigates action label-based random human motion prediction and proposes two memory pools to capture action transitions and action features.
Stop Learning it all to Mitigate Visual Hallucination, Focus on the Hallucination Target.
Dokyoon Yoon (SIONIC AI), Woomyoung Park
RecognitionOptimizationTransformerLarge Language ModelReinforcement LearningImageMultimodality
🎯 What it does: This paper proposes a target learning method TL-DPO for multimodal large language models (MLLMs) to significantly reduce the hallucination problem present in visual information, focusing on correcting erroneous information in specific image regions and response segments.
Stop Walking in Circles! Bailing Out Early in Projected Gradient Descent
Philip Doldo (Booz Allen Hamilton), Edward Raff (Booz Allen Hamilton)
OptimizationComputational EfficiencyAdversarial AttackImage
🎯 What it does: By detecting cycles during the PGD iteration process, an early stopping mechanism for PGD has been implemented for the first time, significantly reducing the computational cost of adversarial robustness evaluation.
STOP: Integrated Spatial-Temporal Dynamic Prompting for Video Understanding
Zichen Liu (Peking University), Jiahuan Zhou (Renmin University of China)
ClassificationRecognitionRetrievalConvolutional Neural NetworkPrompt EngineeringContrastive LearningVideoText
🎯 What it does: This paper proposes an integrated Space-Time Dynamic Prompt (STOP) method, which achieves parameter-efficient fine-tuning for video tasks by incorporating learnable intra-frame spatial prompts and inter-frame temporal prompts into the CLIP pre-trained model.
StoryGPT-V: Large Language Models as Consistent Story Visualizers
Xiaoqian Shen (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
GenerationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: The StoryGPT-V model is proposed, which combines LLM and character-aware Latent Diffusion to achieve story consistency visualization based on co-referential text.
STPro: Spatial and Temporal Progressive Learning for Weakly Supervised Spatio-Temporal Grounding
Aaryan Garg (Birla Institute of Technology and Science Pilani), Yogesh S Rawat (University of Central Florida)
Object DetectionRecurrent Neural NetworkTransformerLarge Language ModelContrastive LearningVideo
🎯 What it does: A weakly supervised spatiotemporal video localization method is proposed, utilizing a vision-language foundation model to locate targets in videos.
StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text
Roberto Henschel (Moonvalley), Humphrey Shi (Picsart AI Research)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkVideoText
🎯 What it does: A self-regressive text-to-long-video generation framework named StreamingT2V is proposed, capable of continuously generating high-quality, motion-rich, and temporally consistent videos of up to two minutes.
StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models
Yunzhi Yan (Zhejiang University), Sida Peng (Zhejiang University)
GenerationData SynthesisAutonomous DrivingDiffusion modelGaussian SplattingVideoPoint Cloud
🎯 What it does: A controllable video diffusion model called StreetCrafter has been developed, which synthesizes street view images using pixel-level conditions from LiDAR and supports editing of dynamic scenes.
Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget
Vikash Sehwag (Sony AI), Lingjuan Lyu (Sony AI)
GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelImageTextMultimodality
🎯 What it does: Trained a text-to-image diffusion transformer model with 1.16B parameters on an extremely low budget (only $1,890), achieving generation quality comparable to large models;
Structure from Collision
Takuhiro Kaneko (NTT Corporation)
RestorationData SynthesisNeural Radiance FieldVideoPhysics Related
🎯 What it does: This paper proposes a model named SfC-NeRF to recover the internal structure of objects from collision videos.
Structure-Aware Correspondence Learning for Relative Pose Estimation
Yihan Chen (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
Pose EstimationTransformerImage
🎯 What it does: A structure-aware correspondence learning method is proposed to estimate relative poses on unseen objects, directly regressing 3D-3D correspondence points without explicit feature matching.
Structure-from-Motion with a Non-Parametric Camera Model
Yihan Wang (École Polytechnique Fédérale de Lausanne), Viktor Larsson (Lund University)
Pose EstimationOptimizationSimultaneous Localization and MappingImage
🎯 What it does: This paper presents an incremental structure from motion (SfM) pipeline named GenSfM, which achieves joint reconstruction and calibration of cameras with arbitrary viewpoint distortion using an adaptive non-parametric camera model.
Structured 3D Latents for Scalable and Versatile 3D Generation
Jianfeng Xiang (Tsinghua University), Jiaolong Yang (Microsoft Research)
GenerationData SynthesisTransformerRectified FlowNeural Radiance FieldPoint CloudMesh
🎯 What it does: A unified Structured LATent (SLAT) representation is proposed, utilizing sparse 3D meshes and multi-view visual features, capable of generating various high-quality 3D assets (Radiance Field, 3D Gaussian, mesh) from text or image prompts, and supports editable and multi-format outputs.
Style Evolving along Chain-of-Thought for Unknown-Domain Object Detection
Zihao Zhang (Tianjin University), Yahong Han (Tianjin University)
Object DetectionDomain AdaptationAutonomous DrivingContrastive LearningImageChain-of-Thought
🎯 What it does: A style evolution method based on chain thinking is proposed, which gradually generates and transfers diverse styles using multi-layer text prompts from words to phrases to complete sentences, achieving single-domain generalization for object detection.
Style Quantization for Data-Efficient GAN Training
Jian Wang (Sichuan University), Jiancheng Lv (Sichuan University)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the Style Quantization GAN (SQ-GAN), which achieves style space quantization by mapping intermediate latent variables to a learnable discrete codebook, significantly improving GAN training and consistency regularization effects under limited data conditions.
Style-Editor: Text-driven Object-centric Style Editing
Jihun Park (DGIST), Sunghoon Im (DGIST)
Image TranslationImage HarmonizationContrastive LearningImageText
🎯 What it does: A text-based object-level style editing model called Style-Editor is proposed, which can accurately perform style transfer on target objects in images based on source text and style text provided by the user.
StyleMaster: Stylize Your Video with Artistic Generation and Translation
Zixuan Ye (Hong Kong University of Science and Technology), Wenhan Luo (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningVideoText
🎯 What it does: StyleMaster achieves high-quality video style transfer and generation by simultaneously utilizing both global and local style information.
StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer
Ruojun Xu (Zhejiang University), Zach Cheng
Image TranslationDiffusion modelImage
🎯 What it does: This paper proposes StyleSSP, a training-free diffusion-based style transfer method that utilizes improved sampling starting points to enhance content preservation and prevent style image content leakage.
StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements
Mingkun Lei (Westlake University), Chi Zhang (Westlake University)
Image TranslationGenerationDiffusion modelImageText
🎯 What it does: A complete framework for text-driven style transfer is proposed, achieving a more precise fusion and layout stability of style and text through cross-modal AdaIN, style-based classifier guidance-free (SCFG), and a teacher model.
Subnet-Aware Dynamic Supernet Training for Neural Architecture Search
Jeimin Jeon (Yonsei University), Bumsub Ham (Yonsei University)
Neural Architecture SearchImage
🎯 What it does: This paper proposes an N-shot NAS supernetwork training framework that dynamically adjusts the training strategy based on subnet characteristics, addressing the issues of low-complexity subnet training bias and momentum noise.
Subspace Constraint and Contribution Estimation for Heterogeneous Federated Learning
Xiangtao Zhang (University of Electronic Science and Technology of China), Le Zhang (University of Electronic Science and Technology of China)
Federated LearningKnowledge DistillationImage
🎯 What it does: The FedSCE method is proposed to address the issues of local overfitting and aggregation bias in Heterogeneous Federated Learning by using low-dimensional subspace constraints at specific layers of the local model and contribution estimation based on the distance of auxiliary model parameters/feature space updates.
Sufficient Invariant Learning for Distribution Shift
Taero Kim (Yonsei University), Kyungwoo Song (Yonsei University)
Domain AdaptationOptimizationContrastive LearningImageBenchmark
🎯 What it does: This paper proposes the Sufficient Invariant Learning (SIL) framework and the Adaptive Sharpness-aware Group Distributionally Robust Optimization (ASGDRO) algorithm, aimed at learning sufficiently many causally invariant features to enhance model robustness when there is distribution shift between training and testing environments.
SUM Parts: Benchmarking Part-Level Semantic Segmentation of Urban Meshes
Weixiao Gao (Delft University of Technology), Hugo Ledoux (Delft University of Technology)
SegmentationMeshBenchmark
🎯 What it does: A large-scale urban texture grid dataset called SUM Parts has been constructed, and fine-grained semantic annotation tools at the face and texture levels are provided.
SuperLightNet: Lightweight Parameter Aggregation Network for Multimodal Brain Tumor Segmentation
Feng Yu (Wuhan Textile University), Minghua Jiang (Wuhan Textile University)
SegmentationConvolutional Neural NetworkMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A super lightweight multimodal brain tumor segmentation network called SuperLightNet is proposed, aiming to significantly reduce the number of parameters and computational cost while maintaining segmentation accuracy.
SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization
Yi Du (University at Buffalo), Chen Wang (University at Buffalo)
Diffusion modelPoint CloudBenchmark
🎯 What it does: A unified diffusion model called SuperPC is proposed, capable of simultaneously completing four tasks: point cloud completion, upsampling, denoising, and coloring.
Supervising Sound Localization by In-the-wild Egomotion
Anna Min (Tsinghua University), Andrew Owens (University of Michigan)
ClassificationRecognitionConvolutional Neural NetworkVideoAudio
🎯 What it does: Utilize self-motion information from videos as weak supervision to train a binaural sound source direction localization model.
SURGEON: Memory-Adaptive Fully Test-Time Adaptation via Dynamic Activation Sparsity
Ke Ma (Northwestern Polytechnical University), Yunhao Liu (Tsinghua University)
SegmentationDomain AdaptationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A method named SURGEON is proposed to significantly reduce memory consumption during the Full Test-Time Adaptation (FTTA) process while maintaining or improving adaptation performance.
SVDC: Consistent Direct Time-of-Flight Video Depth Completion with Frequency Selective Fusion
Xuan Zhu (Huazhong University of Science and Technology), Xin Yang (Huazhong University of Science and Technology)
RestorationDepth EstimationConvolutional Neural NetworkTransformerOptical FlowVideo
🎯 What it does: A video depth completion method called SVDC is proposed to address the sparse noisy depth maps generated by lightweight direct ToF (dToF) sensors for mobile devices, achieving more complete and coherent depth estimation.
SVFR: A Unified Framework for Generalized Video Face Restoration
Zhiyao Wang (Xiamen University), Rongrong Ji (Xiamen University)
RestorationDiffusion modelContrastive LearningVideo
🎯 What it does: This paper proposes a unified framework SVFR that can simultaneously accomplish three major tasks: video blind face restoration, blemish removal, and coloring;
SVG-IR: Spatially-Varying Gaussian Splatting for Inverse Rendering
Hanxiao Sun (Nankai University), Beibei Wang (Nanjing University)
Gaussian SplattingPoint Cloud
🎯 What it does: Introducing spatially varying Gaussian representations and a physically-based indirect lighting model in the inverse rendering task, achieving higher quality novel view synthesis and relighting.
SVLTA: Benchmarking Vision-Language Temporal Alignment via Synthetic Video Situation
Hao Du (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
Data SynthesisTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper designs and implements the SVLTA (Synthetic Vision‑Language Temporal Alignment) benchmark to systematically evaluate the temporal alignment capabilities of vision-language models, and conducts three types of diagnostic experiments on this benchmark: temporal question answering, distribution shift sensitivity, and alignment adaptability.
SwiftEdit: Lightning Fast Text-Guided Image Editing via One-Step Diffusion
Trong-Tung Nguyen (Qualcomm AI Research), Cuong Pham (Qualcomm AI Research)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelImageText
🎯 What it does: This paper presents SwiftEdit, an instant text-guided image editing framework based on a single-step diffusion model, capable of completing edits in 0.23 seconds.
Symbolic Representation for Any-to-Any Generative Tasks
Jiaqi Chen (Stanford University), Li-jia Li (LiveX AI)
GenerationData SynthesisTransformerLarge Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a symbolic generation task description language called A-Language and a zero-shot inference engine based on pre-trained language models, which directly transforms natural language task descriptions into executable multimodal generation workflows.
SymDPO: Boosting In-Context Learning of Large Multimodal Models with Symbol Demonstration Direct Preference Optimization
Hongrui Jia (National Engineering Research Center for Software Engineering Peking University), Shikun Zhang (National Engineering Research Center for Software Engineering Peking University)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageTextMultimodality
🎯 What it does: A method called Symbolic Direct Preference Optimization (SymDPO) is proposed and implemented, which replaces answers with non-semantic symbols in In-context Learning (ICL) to force the model to associate visual information with symbols, enhancing the understanding and utilization of image content.
Symmetry Strikes Back: From Single-Image Symmetry Detection to 3D Generation
Xiang Li (University of Illinois at Urbana-Champaign), James M. Rehg (University of Illinois at Urbana-Champaign)
GenerationOptimizationTransformerDiffusion modelImage
🎯 What it does: A zero-projection symmetry detection framework named Reflect3D has been developed, which can accurately detect 3D reflective symmetry planes from a single RGB image and integrate symmetry information into the single-image 3D generation process to enhance geometric and texture quality.
Synchronized Video-to-Audio Generation via Mel Quantization-Continuum Decomposition
Juncheng Wang (Hong Kong Polytechnic University), Liefeng Bo (Alibaba Group)
GenerationData SynthesisTransformerVideoAudio
🎯 What it does: This paper proposes an audio generation framework based on Mel Quantization-Continuous Decomposition (Mel-QCD), which splits the mel spectrum into three parts: semantics, energy, and standard deviation. Control signals are generated through a video-to-all-signal predictor, enabling the generation of synchronized audio from silent video.
SyncSDE: A Probabilistic Framework for Diffusion Synchronization
Hyunjun Lee (Seoul National University), Sookwan Han (Seoul National University)
GenerationData SynthesisDiffusion modelImageStochastic Differential Equation
🎯 What it does: A synchronous diffusion model SyncSDE based on a probabilistic framework is proposed to achieve collaborative generation among multiple diffusion trajectories.
SyncVP: Joint Diffusion for Synchronous Multi-Modal Video Prediction
Enrico Pallotta (University of Bonn), Juergen Gall (University of Bonn)
GenerationData SynthesisAutonomous DrivingDiffusion modelVideoMultimodality
🎯 What it does: A SyncVP framework is constructed, using a pre-trained latent diffusion model to synchronously predict future video frames across multimodalities such as RGB and depth.
SynerGen-VL: Towards Synergistic Image Understanding and Generation with Vision Experts and Token Folding
Hao Li (Shanghai AI Laboratory), Jifeng Dai (Tsinghua University)
RecognitionGenerationTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: A unified encoder-free multimodal large language model, SynerGen-VL, is proposed, capable of both image understanding and image generation.
Synergizing Motion and Appearance: Multi-Scale Compensatory Codebooks for Talking Head Video Generation
Shuling Zhao (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerAuto EncoderVideo
🎯 What it does: A framework for joint learning of multi-scale motion and appearance compensation codebooks is designed to enhance the motion accuracy and appearance details of lip-sync video generation.
SynTab-LLaVA: Enhancing Multimodal Table Understanding with Decoupled Synthesis
Bangbang Zhou (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityTabular
🎯 What it does: A low-cost and efficient synthesis framework has been designed that splits table synthesis into two steps: image rendering and question-answer generation, and based on this, a SynTab dataset has been generated with a scale of 636K images and 1.8M QA pairs.
Synthetic Data is an Elegant GIFT for Continual Vision-Language Models
Bin Wu (Wuhan University), Mang Ye (Wuhan University)
GenerationData SynthesisKnowledge DistillationTransformerVision Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: A continuous fine-tuning framework named GIFT is proposed, which achieves knowledge retention and generalization balance for Vision-Language models (such as CLIP) by generating synthetic data while continuously learning new tasks.
Synthetic Prior for Few-Shot Drivable Head Avatar Inversion
Wojciech Zielonka (Max Planck Institute for Intelligent Systems), Timo Bolkart (Google)
GenerationData SynthesisConvolutional Neural NetworkTransformerAuto EncoderGaussian SplattingImage
🎯 What it does: The SynShot method is proposed, which can reconstruct a driveable 3D high-quality facial head avatar using only three images.
Synthetic Visual Genome
Jae Sung Park (University of Washington), Ranjay Krishna
Object DetectionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageMultimodalityGraph
🎯 What it does: By automating the construction of Synthetic Visual Graphs (SVG) and training a multimodal language model ROBIN, large-scale dense scene graph generation and visual relationship reasoning are achieved.
Synthetic-to-Real Self-supervised Robust Depth Estimation via Learning with Motion and Structure Priors
Weilong Yan (National University of Singapore), Robby T. Tan (ASUS Intelligent Cloud Services)
Depth EstimationDomain AdaptationAutonomous DrivingDiffusion modelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A self-supervised robust depth estimation framework is proposed, adapting using motion and structure priors from synthetic to real.
SynthLight: Portrait Relighting with Diffusion Model by Learning to Re-render Synthetic Faces
Sumit Chaturvedi (Yale University), Zhixin Shu (Adobe)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper presents SynthLight, a portrait relighting system based on diffusion models that directly learns to 're-render' portraits under different environmental lighting conditions;
T-CIL: Temperature Scaling using Adversarial Perturbation for Calibration in Class-Incremental Learning
Seong-Hyeon Hwang (Korea Advanced Institute of Science and Technology), Steven Euijong Whang (Korea Advanced Institute of Science and Technology)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper studies how to calibrate the model with temperature scaling in Class-Incremental Learning without an old task validation set.
T-FAKE: Synthesizing Thermal Images for Facial Landmarking
Philipp Flotho (Saarland University), Gabriele Steidl (Technical University of Berlin)
RecognitionData SynthesisPose EstimationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This study achieves precise facial annotation in thermal imaging mode by constructing a large-scale synthetic thermal image dataset and training a heatmap facial keypoint detector.
T2ICount: Enhancing Cross-modal Understanding for Zero-Shot Counting
Yifei Qian (University of Nottingham), Michael P. Pound (University of Nottingham)
RecognitionObject DetectionDiffusion modelImageTextMultimodality
🎯 What it does: A zero-shot text-guided counting framework T2ICount based on a single-step diffusion model is designed and implemented to address the issue of insufficient text sensitivity in traditional methods.
T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation
Lijun Li (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
GenerationSafty and PrivacyLarge Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: The T2ISafety benchmark is proposed for system evaluation of text-to-image models in the three major safety areas of fairness, toxicity, and privacy;
T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving
Changsheng Lv (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
Autonomous DrivingTransformerGraph
🎯 What it does: A unified Traffic Topology Scene Graph (TSG) is proposed for traffic scene understanding in autonomous driving, and the TopoFormer model is introduced to achieve one-stage TSG generation.
T2V-CompBench: A Comprehensive Benchmark for Compositional Text-to-video Generation
Kaiyue Sun (University of Hong Kong), Xihui Liu (University of Hong Kong)
Object DetectionObject TrackingGenerationTransformerLarge Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper presents T2V-CompBench, a text-to-video composition generation benchmark containing 1,400 diverse text prompts, and designs evaluation metrics for seven composition categories.
TacoDepth: Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion
Yiran Wang (Nanyang Technological University), Guosheng Lin (Nanyang Technological University)
Depth EstimationAutonomous DrivingGraph Neural NetworkMultimodalityPoint Cloud
🎯 What it does: This paper presents TacoDepth, a one-stage fusion radar-camera depth estimation model.
TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning
Seungmin Baek (Ewha Womans University), Dongbo Min (Ewha Womans University)
SegmentationTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes a Task-Adaptive Dynamic Transformer (TADFormer) for multi-task learning, utilizing learnable task prompts and dynamic task filters to achieve efficient fine-grained feature extraction for multiple tasks without significantly increasing the number of trainable parameters.
TAET: Two-Stage Adversarial Equalization Training on Long-Tailed Distributions
Wang Yu-Hang (Hefei University of Technology), Jian Liu (Hefei University of Technology)
ClassificationAdversarial AttackGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A two-stage adversarial equalization training (TAET) framework is proposed to enhance the model's adversarial robustness and balanced accuracy under long-tail distributions.
TAGA: Self-supervised Learning for Template-free Animatable Gaussian Articulated Model
Zhichao Zhai (Zhejiang University), Jun Xiao (Zhejiang University)
GenerationPose EstimationAnomaly DetectionOptimizationGaussian SplattingVideo
🎯 What it does: This paper proposes a self-supervised, template-free framework for animatable model reconstruction of humans/animals, called TAGA. It utilizes explicit 3D Gaussian representations and skinning driven by skeletal structures to simultaneously learn geometry and skinning weights from sparse pose videos.
TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection
Yoon Gyo Jung (Northeastern University), Octavia Camps (Northeastern University)
Anomaly DetectionImage
🎯 What it does: In the industrial anomaly detection scenario where the training data exhibits both long-tail distribution and noise, a memory-based anomaly detection framework called TailedCore is proposed. It predicts the class scale of samples through TailSampler, thereby achieving incremental memory by sampling only a few minority class samples.
Take the Bull by the Horns: Learning to Segment Hard Samples
Yuan Guo (Tianjin University), Yuping Duan (Beijing Normal University)
SegmentationTransformerDiffusion modelBiomedical DataUltrasound
🎯 What it does: This paper proposes the L2S framework, which can automatically identify hard samples in medical images and use a specialized segmentation network for precise segmentation.
Taming Teacher Forcing for Masked Autoregressive Video Generation
Deyu Zhou (Hong Kong University of Science and Technology), Xiangyu Zhang (StepFun)
GenerationData SynthesisTransformerVideo
🎯 What it does: A hybrid video generation framework MAGI is designed, integrating mask modeling and causal modeling to achieve high-quality and scalable video generation.
Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs
Yingji Zhong (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisOptimizationDiffusion modelGaussian SplattingVideoPoint Cloud
🎯 What it does: In sparse input 3D Gaussian Splatting (3DGS) scenes, multi-view sequences are generated using video diffusion models, addressing extrapolation and occlusion issues through scene grounding.
TAMT: Temporal-Aware Model Tuning for Cross-Domain Few-Shot Action Recognition
Yilong Wang (Tianjin University), Qinghua Hu (Tianjin University)
RecognitionDomain AdaptationTransformerSupervised Fine-TuningVideo
🎯 What it does: Proposes the Temporal-Aware Model Tuning (TAMT), a cross-domain few-shot action recognition framework based on separate training.
TANGO: Training-free Embodied AI Agents for Open-world Tasks
Filippo Ziliotto (University of Padova), Lamberto Ballan (University of Padova)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper presents TANGO, a training-agnostic neural symbolic framework based on large language model planning, pre-trained visual and navigation modules, designed to perform various embedded AI tasks.
TaoAvatar: Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting
Jianchuan Chen (Alibaba Group), Chengfei Lv (Alibaba Group)
GenerationKnowledge DistillationGaussian SplattingMesh
🎯 What it does: Using 3D Gaussian Splatting and a personalized SMPLX++ clothing extension template, we construct a fully-renderable Talking Avatar that can be rendered in real-time, achieving high-quality details through teacher-student distillation and lightweight Blend Shapes.
TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models
Xin Wang (Fudan University), Xingjun Ma (Fudan University)
OptimizationAdversarial AttackTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a method called Test-time Adversarial Prompt Tuning (TAPT) to enhance robustness in zero-shot inference of visual-language models like CLIP.
Targeted Forgetting of Image Subgroups in CLIP Models
Zeliang Zhang (University of Rochester), Chenliang Xu (University of Rochester)
ClassificationKnowledge DistillationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A three-stage subgroup forgetting method for the CLIP model is proposed, allowing for selective forgetting of certain subsets of images without accessing the original training data.
TAROT: Towards Essentially Domain-Invariant Robustness with Theoretical Justification
Dongyoon Yang (SK Hynix), Yongdai Kim (Seoul National University)
Domain AdaptationAdversarial AttackImage
🎯 What it does: A theoretical framework and algorithm TAROT is proposed, which simultaneously considers adversarial robustness and unsupervised domain adaptation.
Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics
Shibo Zhao (Carnegie Mellon University), Sebastian Scherer (Carnegie Mellon University)
Pose EstimationRobotic IntelligenceRecurrent Neural NetworkMixture of ExpertsSimultaneous Localization and MappingTime Series
🎯 What it does: A foundational model named Tartan IMU is proposed, utilizing over 100 hours of multi-platform IMU data to achieve general pose estimation. On this basis, LoRA fine-tuning and an online adaptive mechanism are added to support real-time localization on various robotic platforms.
Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment
Ziang Yan (Shanghai AI Laboratory), Yi Wang (Shanghai Innovation Institute)
SegmentationRetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningImageVideoTextMultimodality
🎯 What it does: This paper proposes the Task Preference Optimization (TPO) method, which utilizes differentiable task preferences to couple various visual task heads with a multimodal large language model (MLLM), enhancing the model's capabilities in fine-grained visual understanding and dialogue.
Task Singular Vectors: Reducing Task Interference in Model Merging
Antonio Andrea Gargiulo (Sapienza University of Rome), Emanuele Rodolà (Sapienza University of Rome)
ClassificationRecognitionTransformerSupervised Fine-TuningImage
🎯 What it does: The TSV-Merge method is proposed to achieve model merging without additional training;