ICCV 2025 Papers — Page 23
IEEE/CVF International Conference on Computer Vision · 2701 papers
SPD: Shallow Backdoor Protecting Deep Backdoor Against Backdoor Detection
Shunjie Yuan (Xidian University), Robert H. Deng (Singapore Management University)
Adversarial AttackAuto EncoderImage
🎯 What it does: A white-box backdoor attack method called SPD is designed, combining shallow pixel domain triggers with deep frequency domain amplitude triggers to achieve the concealment and defense detection of backdoors.
SpecGuard: Spectral Projection-based Advanced Invisible Watermarking
Inzamamul Alam (Sungkyunkwan University), Khan Muhammad (Sungkyunkwan University)
Image TranslationData SynthesisSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an implicit watermark embedding method called SpecGuard, which embeds information into the high-frequency spectrum domain without compromising image quality and achieves robust extraction.
Spectral Image Tokenizer
Carlos Esteves (Google Research), Ameesh Makadia (Google Research)
GenerationData SynthesisTransformerGenerative Adversarial NetworkImageText
🎯 What it does: This paper proposes the Spectral Image Tokenizer (SIT), which first applies Discrete Wavelet Transform (DWT) to the image and then encodes it into a sequence of discrete tokens using a transformer, supporting multi-scale decoding from coarse to fine.
Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating
Lilika Makabe (Osaka University), Yasuyuki Matsushita (Osaka University)
Image
🎯 What it does: A method for estimating the spectral sensitivity of a camera using an uncalibrated diffraction grating single-lens capture is proposed.
SpectralAR: Spectral Autoregressive Visual Generation
Yuanhui Huang (Tsinghua University), Jiwen Lu (Tsinghua University)
GenerationTransformerLarge Language ModelImage
🎯 What it does: Proposes SpectralAR, which implements autoregressive visual generation from a spectral perspective, converting images into 1D spectral sequences for hierarchical coarse-to-fine generation.
Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation
Pierre-André Brousseau (Université de Montréal), Sébastien Roy (Université de Montréal)
Depth EstimationImageVideo
🎯 What it does: A differentiable spherical epipolar correction model is proposed, combined with a self-supervised deep network to achieve absolute depth estimation from two frames of monocular video.
SpikeDiff: Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams
Siqi Yang (Peking University), Boxin Shi (Peking University)
RestorationGenerationDiffusion modelVideo
🎯 What it does: Reconstructing sub-millisecond temporal color video in zero-shot scenarios using pre-trained diffusion models
SpikePack: Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility
Guobin Shen (Chinese Academy of Sciences Institute of Automation), Yi Zeng (Chinese Academy of Sciences Institute of Automation)
Object DetectionSegmentationComputational EfficiencySpiking Neural NetworkImage
🎯 What it does: This paper proposes a synaptic model named SpikePack, which significantly reduces information loss during the synaptic transmission process while maintaining membrane potential reset and leakage integration characteristics.
SpiLiFormer: Enhancing Spiking Transformers with Lateral Inhibition
Zeqi Zheng (Zhejiang University), Yaochu Jin (Westlake University)
ClassificationSpiking Neural NetworkTransformerImage
🎯 What it does: A spike neural network based on Transformer, called SpiLiFormer, is proposed, which improves spike attention using a brain-inspired lateral inhibition mechanism to enhance image classification performance.
SpinMeRound: Consistent Multi-View Identity Generation Using Diffusion Models
Stathis Galanakis (Imperial College London), Stefanos Zafeiriou (Imperial College London)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Introducing SpinMeRound—a multi-view diffusion model that utilizes identity embeddings and multi-view conditions to generate high-fidelity full-head images and corresponding surface normals, achieving 3D consistent face synthesis from different angles.
SplArt: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting
Shengjie Lin (Toyota Technological Institute at Chicago), Matthew R. Walter (Toyota Technological Institute at Chicago)
RestorationSegmentationPose EstimationGaussian SplattingImage
🎯 What it does: This paper proposes a self-supervised two-state RGB image recovery method called SPLART, which simultaneously reconstructs and infers the motion model of movable parts using 3D Gaussian Splatting, achieving real-time and realistic viewpoint synthesis.
Splat-based 3D Scene Reconstruction with Extreme Motion-blur
Hyeonjoong Jang (Korea Advanced Institute of Science and Technology), Min H. Kim (Korea Advanced Institute of Science and Technology)
RestorationPose EstimationDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingOptical FlowImagePoint Cloud
🎯 What it does: A robust 3D scene reconstruction method based on RGB-D input is proposed, capable of simultaneously achieving camera pose estimation, point cloud refinement, and image deblurring under extreme motion blur.
Splat-LOAM: Gaussian Splatting LiDAR Odometry and Mapping
Emanuele Giacomini (Sapienza University of Rome), Martin R. Oswald (University of Amsterdam)
Autonomous DrivingOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A LiDAR odometry and mapping pipeline called Splat-LOAM based on 2D Gaussian primitives is proposed.
SplatTalk: 3D VQA with Gaussian Splatting
Anh Thai (Georgia Institute of Technology), Thomas Funkhouser (Google DeepMind)
RecognitionGenerationRetrievalTransformerLarge Language ModelVision Language ModelAuto EncoderGaussian SplattingImageMultimodalityPoint Cloud
🎯 What it does: Using multi-view RGB images, we propose SplatTalk, which generates a 3D semantic field by projecting language features extracted from a vision-language model into a 3D Gaussian distribution, allowing for zero-shot 3D visual question answering directly inputting into a large language model;
Split-and-Combine: Enhancing Style Augmentation for Single Domain Generalization
Zhen Zhang (Anhui Agricultural University), Lichuan Gu (Anhui Agricultural University)
Data SynthesisDomain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: A strategy named 'Split-And-Combine' (SAC) is proposed to enhance the style augmentation capability in single-domain generalization. This method improves the diversity of training data by segmenting images into multiple patches and performing style augmentation independently on each patch.
SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement
Liwen Xiao (Huazhong University of Science and Technology), Wei Li (Nanyang Technological University)
Autonomous DrivingTime Series
🎯 What it does: A multi-agent trajectory refinement framework SRefiner based on soft junction topology is proposed, which captures the spatiotemporal topological relationship between trajectories and lanes through soft junctions, and continuously updates the topological information in multiple iterations to improve prediction accuracy.
SSVQ: Unleashing the Potential of Vector Quantization with Sign-Splitting
Shuaiting Li (Zhejiang University), Kejie Huang (Zhejiang University)
Object DetectionSegmentationGenerationCompressionSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Improving the limitations of traditional vector quantization during the fine-tuning phase, a learnable symbol-split vector quantization (SSVQ) is proposed, allowing each quantization weight to independently follow gradient updates;
St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World
Haiwen Feng (Max Planck Institute for Intelligent Systems), Angjoo Kanazawa (University of California Berkeley)
Object TrackingDepth EstimationTransformerSimultaneous Localization and MappingVideoPoint CloudBenchmark
🎯 What it does: An end-to-end framework called St4RTrack is proposed for simultaneous 3D reconstruction and dynamic video point tracking, capable of obtaining 3D point clouds and trajectories in world coordinates from monocular videos.
Stable Diffusion Models are Secretly Good at Visual In-Context Learning
Trevine Oorloff (Apple), Reza Ardekani (Apple)
Object DetectionSegmentationTransformerDiffusion modelImage
🎯 What it does: This study demonstrates that the directly reusable Stable Diffusion model can achieve visual context learning during the inference phase without additional training or data.
Stable Score Distillation
Haiming Zhu (Singapore Management University), Shengfeng He (Singapore Management University)
GenerationData SynthesisOptimizationTransformerPrompt EngineeringDiffusion modelScore-based ModelImagePoint Cloud
🎯 What it does: A Stable Score Distillation (SSD) method is proposed for text-guided image and 3D scene editing, which achieves more realistic editing effects while maintaining the original content structure.
Stable Virtual Camera: Generative View Synthesis with Diffusion Models
Jensen Zhou, Varun Jampani (Stability AI)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: A general diffusion model SEVA has been developed to generate new views from any number of input views, supporting large viewpoint changes and temporally smooth videos.
Stable-Sim2Real: Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion
Mutian Xu (Chinese University of Hong Kong Shenzhen), Xiaoguang Han (Chinese University of Hong Kong Shenzhen)
SegmentationGenerationData SynthesisDiffusion modelPoint Cloud
🎯 What it does: A 3D Sim2Real method called Stable-Sim2Real is proposed, which is based on a two-stage diffusion model to generate realistic 3D point clouds from synthetic depth maps.
StableCodec: Taming One-Step Diffusion for Extreme Image Compression
Tianyu Zhang (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
CompressionDiffusion modelAuto EncoderImage
🎯 What it does: A novel ultra-low bitrate image compression method called StableCodec based on first-order diffusion is proposed, capable of generating highly realistic and high-fidelity images in the range of 0.005–0.05 bpp.
StableDepth: Scene-Consistent and Scale-Invariant Monocular Depth
Zheng Zhang (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Depth EstimationTransformerDiffusion modelVideo
🎯 What it does: StableDepth is a monocular depth estimation method that achieves scene-consistent and scale-invariant depth prediction, supporting online inference.
Staining and Locking Computer Vision Models Without Retraining
Oliver J. Sutton (Synoptix Ltd.), Ivan Y. Tyukin (King's College London)
Object DetectionData SynthesisSafty and PrivacyConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: An algorithm is proposed for embedding watermarks (stain) and locks in pre-trained visual models without the need for retraining, aimed at protecting the intellectual property of the model.
STaR: Seamless Spatial-Temporal Aware Motion Retargeting with Penetration and Consistency Constraints
Xiaohang Yang (Queen Mary University of London), Shanxin Yuan (Queen Mary University of London)
TransformerPoint CloudMesh
🎯 What it does: STaR proposes an end-to-end spatiotemporal motion remapping framework that can generate complete, geometrically reasonable, and temporally coherent motion for target characters from the motion sequences of source characters.
STAR: Spatial-Temporal Augmentation with Text-to-Video Models for Real-World Video Super-Resolution
Rui Xie (Nanjing University), Ying Tai (Nanjing University)
RestorationSuper ResolutionDiffusion modelVideoText
🎯 What it does: The STAR framework is proposed, utilizing a text-to-video diffusion model along with local information enhancement and dynamic frequency loss to achieve super-resolution of real scene videos.
Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images
Qi Xun Yeo (National University of Singapore), Gim Hee Lee (National University of Singapore)
Object DetectionSegmentationGenerationRetrievalGraph Neural NetworkSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A framework for generating 3D semantic scene graphs based on multi-view RGB images is proposed, which mainly includes modules for initializing node features using segmentation masks, improving edge features based on residual neighborhood graph convolutional networks, and confidence re-labeling using statistical priors.
STD-GS: Exploring Frame-Event Interaction for SpatioTemporal-Disentangled Gaussian Splatting to Reconstruct High-Dynamic Scene
Hanyu Zhou (Huazhong University of Science and Technology), Gim Hee Lee (National University of Singapore)
RestorationSegmentationGenerationGaussian SplattingMultimodality
🎯 What it does: A spatiotemporal decoupling framework for high dynamic scene reconstruction, STD-GS, is proposed based on event camera assistance. It first decomposes dynamic scenes into background and dynamic objects, then fuses spatial color and temporal deformation on 4D Gaussian points, ultimately achieving temporally continuous rendering of high dynamic scenes.
STDDNet: Harnessing Mamba for Video Polyp Segmentation via Spatial-aligned Temporal Modeling and Discriminative Dynamic Representation Learning
Guilian Chen (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationRepresentation LearningConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes STDDNet, a video polyp segmentation network that combines spatially aligned temporal modeling with discriminative dynamic representation learning.
StealthAttack: Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions
Bo-Hsu Ke (National Yang Ming Chiao Tung University), Wei-Chen Chiu (National Yang Ming Chiao Tung University)
Adversarial AttackGaussian SplattingPoint Cloud
🎯 What it does: This paper investigates the security vulnerabilities of 3D Gaussian Splatting (3DGS) and proposes a density-guided poisoning attack that can inject artifact objects into the target view while maintaining the rendering quality of other views without significant impact.
Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment
Qingqian Yang (Shanghai University of Electric Power), Haibing Guan (Shanghai Jiao Tong University)
Federated LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposed and implemented an adaptive layer-wise gradient alignment attack (LGA) to inject covert backdoors in federated learning;
Steering Guidance for Personalized Text-to-Image Diffusion Models
Sunghyun Park (Qualcomm AI Research), Sungrack Yun (Qualcomm AI Research)
GenerationData SynthesisSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: The Personalization Guidance (PG) method is proposed, which forms a 'weak model' by interpolating the weights of the pre-trained model and the fine-tuned model during the inference phase, thereby enhancing the adaptability to the target concept while maintaining text consistency.
SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering
Byeongjun Park (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: This paper proposes a zero-shot inference guidance method named SteerX, which unifies video generation and scene reconstruction. It dynamically adjusts the sampling distribution during the generation process using a geometric reward function, achieving high-quality 3D and 4D scene generation under no-camera conditions.
STEP-DETR: Advancing DETR-based Semi-Supervised Object Detection with Super Teacher and Pseudo-Label Guided Text Queries
Tahira Shehzadi (Deutsche Forschungsgemeinschaft), Muhammad Zeshan Afzal (Deutsche Forschungsgemeinschaft)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: This paper proposes STEP-DETR, a semi-supervised object detection framework based on DETR, which incorporates Super Teacher, Pseudo-Label Text Queries, Denoising Text Guided Object Queries, and Query Refinement to enhance pseudo-label quality, balance the detection of common and rare categories, distinguish between foreground/background, and eliminate redundant queries.
Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation
Yong Liu (Tsinghua University), Yansong Tang (Tsinghua University)
SegmentationConvolutional Neural NetworkVision Language ModelImageBenchmark
🎯 What it does: A new benchmark called OpenBench is proposed along with the OVSNet model, which enhances open vocabulary segmentation performance.
Stereo Any Video: Temporally Consistent Stereo Matching
Junpeng Jing (Imperial College London), Krystian Mikolajczyk (Imperial College London)
Image TranslationDepth EstimationConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: Designed and implemented the Stereo Any Video framework for dynamic video stereo matching without relying on camera pose or optical flow information.
STI-Bench: Are MLLMs Ready for Precise Spatial-Temporal World Understanding?
Yun Li (Shanghai Jiao Tong University), Bo Zhao (Shanghai Jiao Tong University)
Autonomous DrivingTransformerLarge Language ModelPrompt EngineeringVideoMultimodalityBenchmark
🎯 What it does: This paper proposes and implements STI-Bench, a benchmark based on 300 real video segments and 2,000 precisely quantified questions, designed to evaluate the capabilities of multimodal large language models in spatial-temporal understanding.
STIV: Scalable Text and Image Conditioned Video Generation
Zongyu Lin (Apple), Kai-Wei Chang (University of California)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: STIV integrates a variable number of image conditions with text conditions in the Diffusion Transformer (DiT) using frame replacement technology, achieving a unified video generation model capable of performing both text-to-video (T2V) and text+image-to-video (TI2V) tasks, and supports autoregressive long video generation.
Stochastic Gradient Estimation for Higher-Order Differentiable Rendering
Zican Wang (University College London), Tobias Ritschel (University College London)
OptimizationConvolutional Neural NetworkMesh
🎯 What it does: This study investigates how to use importance sampling to estimate the second derivatives (Hessian and Hessian-Vector products) of a renderer and applies it to inverse rendering optimization.
Stochastic Interpolants for Revealing Stylistic Flows across the History of Art
Pingchuan Ma (CompVis at Ludwig Maximilian University of Munich), Björn Ommer (CompVis at Ludwig Maximilian University of Munich)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelImage
🎯 What it does: This paper proposes an unsupervised generative framework that utilizes stochastic interpolation and dual diffusion implicit bridges, capable of mapping a single artwork to different historical style distributions without paired data, thereby revealing the evolution of artistic styles over time.
StochasticSplats: Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting
Shakiba Kheradmand (Google DeepMind), Andrea Tagliasacchi (Simon Fraser University)
Gaussian SplattingPoint Cloud
🎯 What it does: A non-sorting 3D Gaussian fog rendering method is proposed, utilizing stochastic transparency (Monte Carlo estimation) to replace traditional depth sorting and alpha blending, achieving volumetric rendering.
StolenLoRA: Exploring LoRA Extraction Attacks via Synthetic Data
Yixu Wang (Fudan University), Xingjun Ma (Shanghai Artificial Intelligence Laboratory)
Data SynthesisAdversarial AttackTransformerLarge Language ModelDiffusion modelImage
🎯 What it does: Research on model extraction attacks for LoRA adaptation models and the proposal of the StolenLoRA method.
Straighten Viscous Rectified Flow via Noise Optimization
Jimin Dai (Nanjing University of Science and Technology), Lei Luo (Xidian University)
GenerationOptimizationDiffusion modelRectified FlowAuto EncoderImage
🎯 What it does: This paper proposes the VRFNO (Straighten Viscous Rectified Flow via Noise Optimization) framework, which jointly trains a VAE-style encoder and a neural velocity field. By reparameterizing noise, it achieves optimized coupling, resulting in smoother straight-line approximations of probabilistic flow trajectories, enabling high-quality image generation in one or few steps.
StrandHead: Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors
Xiaokun Sun (Nanjing University), Zhenyu Zhang (Nanjing University)
GenerationDiffusion modelScore-based ModelTextMesh
🎯 What it does: A method for generating 3D head avatars based on text is proposed, and the generation and editing of discrete hair strands are implemented.
StreamDiffusion: A Pipeline-level Solution for Real-Time Interactive Generation
Akio Kodaira (University of California Berkeley), Kurt Keutzer (University of California Berkeley)
GenerationComputational EfficiencyDiffusion modelImageVideo
🎯 What it does: An end-to-end real-time streaming diffusion pipeline, StreamDiffusion, is proposed to achieve high-throughput interactive image generation.
StreamGS: Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams
Yang Li (Microsoft Research), Yan Lu (Microsoft Research)
RestorationDomain AdaptationComputational EfficiencyGaussian SplattingImageVideo
🎯 What it does: A method for online, camera pose information-free, and generalizable 3D Gaussian Splatting reconstruction called StreamGS is proposed, which can generate high-quality 3D scenes frame by frame in real-time from a continuous stream of uncalibrated images and perform novel view synthesis.
Streaming VideoLLMs for Real-Time Procedural Video Understanding
Dibyadip Chatterjee (Meta Reality Labs), Fadime Sener (Meta Reality Labs)
RecognitionCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: ProVideLLM is proposed, a real-time streaming program video understanding framework that utilizes multimodal caching to compress long-term memory into text tokens while retaining short-term details as visual tokens.
Streamlining Image Editing with Layered Diffusion Brushes
Peyman Gholami (University of British Columbia), Robert Xiao (University of British Columbia)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper proposes Layered Diffusion Brushes (LDB), a training-free, interactive, hierarchical image editing framework based on diffusion models, which supports object addition and deletion, attribute modification, and style transformation in any region while preserving the background.
StreamMind: Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition
Xin Ding (University of Science and Technology of China), Ting Cao (Institute for AI Industry Research)
RecognitionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVideo
🎯 What it does: Proposes the STREAMMIND framework, which utilizes event-driven LLM calls to achieve a real-time video dialogue system with a full frame rate of 100fps.
Street Gaussians without 3D Object Tracker
Ruida Zhang (Tsinghua University), Gim Hee Lee (National University of Singapore)
Object TrackingData SynthesisAutonomous DrivingGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Utilizing a 2D foundational model for vehicle tracking, and learning point motion in an implicit feature space to achieve self-correction of 3D tracking errors, completing dynamic 3D scene reconstruction and new perspective synthesis in street scenes.
Stroke2Sketch: Harnessing Stroke Attributes for Training-Free Sketch Generation
Rui Yang (Huaqiao University), Shengfeng He (Singapore Management University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper studies an untrained sketch generation framework called Stroke2Sketch, which can accurately transfer the style of content images using the stroke attributes of reference sketches while maintaining the integrity of the semantic structure.
Stronger, Steadier & Superior: Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation
Siyu Chen (Jimei University), Jinhe Su (Jimei University)
SegmentationDomain AdaptationAutonomous DrivingTransformerSupervised Fine-TuningImage
🎯 What it does: DepthForge fine-tunes domain generalization semantic segmentation by integrating a pre-trained depth VFM with a visual VFM, utilizing depth-aware learnable tokens.
Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation
Guanyi Qin (National University of Singapore), Yueming Jin (National University of Singapore)
Object DetectionSegmentationConvolutional Neural NetworkVideo
🎯 What it does: We propose OASIS, a memory-based semi-supervised video object segmentation method that utilizes Canny edge priors, a lightweight structural refinement module, and evidence learning to achieve boundary refinement and uncertainty suppression.
Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning
Tan Pan (Fudan University), Yuan Cheng (Fudan University)
ClassificationImage TranslationRestorationSegmentationTransformerContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: This paper proposes a structure-aware 3D medical image self-supervised learning framework SDC2, aimed at learning the semantic differences and consistencies between different anatomical structures.
Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal
Wanchang Yu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
RestorationDiffusion modelImage
🎯 What it does: A shadow removal method for portraits based on diffusion models is proposed, treating shadow removal as a structure-guided filling task.
Structured Policy Optimization: Enhance Large Vision-Language Model via Self-referenced Dialogue
Guohao Sun (Rochester Institute of Technology), Zhiqiang Tao (Rochester Institute of Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageVideoTextMultimodality
🎯 What it does: A Structured Policy Optimization (SPO) method is proposed for preference optimization of large visual-language models (LVLM), enabling the model to simultaneously focus on the visual dependencies of question generation and answer generation in multimodal dialogues.
StruMamba3D: Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning
Chuxin Wang (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Representation LearningConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: Proposed StruMamba3D based on Structured State Space Model (SSM) for self-supervised point cloud representation learning.
StyleKeeper: Prevent Content Leakage using Negative Visual Query Guidance
Jaeseok Jeong (Yonsei University), Youngjung Uh (Yonsei University)
GenerationDiffusion modelImage
🎯 What it does: Proposes the StyleKeeper method to generate images under visual style prompts, avoiding content leakage.
StyleMotif: Multi-Modal Motion Stylization using Style-Content Cross Fusion
Ziyu Guo (Chinese University of Hong Kong), Mubbasir Kapadia (Roblox)
GenerationRetrievalDiffusion modelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: This paper proposes STYLEMOTIF, a multimodal motion stylization framework based on latent diffusion models, which can simultaneously control motion content and style inputs from various modalities such as text, images, videos, and audio.
StyleSRN: Scene Text Image Super-Resolution with Text Style Embedding
Shengrong Yuan (Hunan Normal University), Nong Sang (Huazhong University of Science and Technology)
RecognitionRestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A style-embedded scene text image super-resolution network, StyleSRN, is proposed to simultaneously restore text structure and style.
Stylized-Face: A Million-level Stylized Face Dataset for Face Recognition
Zhengyuan Peng (Shanghai Jiao Tong University), Lizhuang Ma (East China Normal University)
RecognitionGenerationDiffusion modelImageBenchmark
🎯 What it does: A Stylized-Face dataset (4.6M images, 62k IDs, 80 styles) was constructed and three sets of evaluation benchmarks were proposed for studying generative style face recognition.
SU-RGS: Relightable 3D Gaussian Splatting from Sparse Views under Unconstrained Illuminations
Qi Zhang (Tianjin University), Wei Feng (Tianjin University)
RestorationGenerationGaussian SplattingPoint Cloud
🎯 What it does: A method named SU-RGS is proposed, capable of performing 3D Gaussian Splatting inverse rendering under conditions of limited viewpoints and unconstrained lighting, outputting new viewpoints, surface materials, geometry, and re-illuminated results.
SUB: Benchmarking CBM Generalization via Synthetic Attribute Substitutions
Jessica Bader (Technical University of Munich), Zeynep Akata (Technical University of Munich)
RecognitionGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: This paper proposes the SUB (Synthetic Attribute Substitutions) benchmark dataset to evaluate the generalization ability of Concept Bottleneck Models (CBM) and Visual Language Models (VLM) when faced with fine-grained attribute substitutions, and generates high-quality attribute substitution images through a new method called Tied Diffusion Guidance (TDG);
Subjective Camera 1.0: Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion
Haoyang Chen (Wuhan University), Zhixiang Wang (CyberAgent AI Lab)
GenerationOptimizationDiffusion modelImageText
🎯 What it does: Introducing the concept of a subjective camera, utilizing text and sequence sketches in an untrained manner to reconstruct real scenes guided by sequence-aware sketches through a diffusion model.
SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models
Kien Nguyen (Qualcomm AI Research), Cuong Pham (Posts and Telecommunications Institute of Technology)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: The SuMa method is proposed for robust and effective concept ablation of narrow concepts in text-to-image diffusion models.
SummDiff: Generative Modeling of Video Summarization with Diffusion
Kwanseok Kim (Seoul National University), Joonseok Lee (Seoul National University)
GenerationTransformerDiffusion modelVideo
🎯 What it does: A generative video summarization method based on diffusion models, SummDiff, is proposed, which can learn and generate various summaries that align with human perspectives.
Super Resolved Imaging with Adaptive Optics
Robin Swanson (University of Toronto), Kiriakos N. Kutulakos (University of Toronto)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: The paper proposes using existing adaptive optics systems to achieve multi-image super-resolution by applying learnable phase distortions to the wavefront, thereby enhancing optical resolution while maintaining AO correction.
Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts
Viet Nguyen (Qualcomm AI Research), Anh Tran (Qualcomm AI Research)
GenerationData SynthesisKnowledge DistillationPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes the NASA method, which directly integrates negative prompts into the cross-attention layer of a single-step diffusion model to achieve controllable suppression of generated images.
Supercharging Floorplan Localization with Semantic Rays
Yuval Grader (Tel Aviv University), Hadar Averbuch-Elor (Cornell University)
Object DetectionPose EstimationDepth EstimationConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A semantic-aware floor plane localization framework is proposed, which jointly estimates depth rays and semantic rays, and constructs a structure-semantic probability volume through coarse-to-fine sampling to localize the camera.
SuperDec: 3D Scene Decomposition with Superquadrics Primitives
Elisabetta Fedele (ETH Zurich), Francis Engelmann (Stanford University)
SegmentationRobotic IntelligenceTransformerPoint Cloud
🎯 What it does: A 3D scene decomposition method based on superquadrics, called SUPERDEC, is designed to decompose point clouds into a small number of superquadrics for a compact and interpretable representation.
SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing
Ming Li (ByteDance Intelligent Creation), Sijie Zhu (ByteDance Intelligent Creation)
Image TranslationGenerationLarge Language ModelDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: This paper proposes SuperEdit for instruction-driven image editing, utilizing VLM (GPT-4o) to correct the alignment of original-edit pairs and further enhance the quality of the supervisory signal through contrastive instructions and triplet loss, thereby improving editing results without adding extra models or pre-training.
SuperEvent: Cross-Modal Learning of Event-based Keypoint Detection for SLAM
Yannick Burkhardt (Technical University of Munich), Stefan Leutenegger (ETH Zurich)
Pose EstimationTransformerSimultaneous Localization and MappingMultimodality
🎯 What it does: A learnable keypoint detection and descriptor called SuperEvent based on event cameras is proposed, achieving stable event feature matching through pseudo-label training, and subsequently integrated into a visual-inertial SLAM framework.
SuperMat: Physically Consistent PBR Material Estimation at Interactive Rates
Yijia Hong (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
GenerationDiffusion modelMesh
🎯 What it does: This paper presents SuperMat, a single-step PBR material decomposition framework based on Stable Diffusion, which can simultaneously generate Albedo, Metallic, and Roughness maps, and achieves efficient material estimation for 3D objects through a UV refinement network.
Superpowering Open-Vocabulary Object Detectors for X-ray Vision
Pablo Garcia-Fernandez (University of Santiago de Compostela), Elisa Ricci (University of Trento)
Object DetectionTransformerLarge Language ModelImageMultimodality
🎯 What it does: This paper proposes RAXO, a training-free cross-modal adaptation framework that can directly transfer existing RGB open-vocabulary object detection (OvOD) models to X-ray images, enabling detection of any user-specified category.
Supervised Exploratory Learning for Long-Tailed Visual Recognition
Zhongquan Jian (Minjiang University), Qingqiang Wu (Xiamen University)
ClassificationRecognitionImage
🎯 What it does: This paper proposes a simple and effective Supervised Exploration Learning (SEL) framework that generates exploratory examples through spatial exploration to improve the decision regions and boundaries in long-tail visual recognition.
SurfaceSplat: Connecting Surface Reconstruction and Gaussian Splatting
Zihui Gao (Zhejiang University), Chunhua Shen (Zhejiang University)
RestorationGenerationGaussian SplattingPoint Cloud
🎯 What it does: Proposes the SurfaceSplat method, achieving surface reconstruction and novel view rendering under sparse perspectives.
SUV: Suppressing Undesired Video Content via Semantic Modulation Based on Text Embeddings
Xiang Lv (China University of Petroleum), Xinyuan Chen (China University of Petroleum)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: A semantic modulation method based on text embedding, called SUV, is proposed for suppressing unwanted content and maintaining temporal consistency in text-driven video editing.
SV4D 2.0: Enhancing Spatio-Temporal Consistency in Multi-View Video Diffusion for High-Quality 4D Generation
Chun-Han Yao (Stability AI), Varun Jampani (Stability AI)
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldVideo
🎯 What it does: This paper presents SV4D 2.0, a multi-view video synthesis network based on diffusion models, which can generate high-quality, spatiotemporally consistent new view videos from monocular videos and further optimize 4D assets.
SVG-Head: Hybrid Surface-Volumetric Gaussians for High-Fidelity Head Reconstruction and Real-Time Editing
Heyi Sun (Tsinghua University), Song-Hai Zhang (Tsinghua University)
GenerationOptimizationGaussian SplattingVideoMesh
🎯 What it does: This paper proposes the Surface-Volumetric Gaussian Head Avatar (SVG-Head), which constructs a real-time editable head avatar model using surface Gaussians bound to the FLAME mesh and volumetric Gaussians that can move freely in voxel space, achieving explicit texturing of surface Gaussians for texture editing.
SViM3D: Stable Video Material Diffusion for Single Image 3D Generation
Andreas Engelhardt (Stability AI), Varun Jampani (Stability AI)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: Developed SViM3D, a multi-view PBR material generation framework based on video diffusion models, capable of predicting multi-view consistent RGB, material parameters, and normals from a single image and camera trajectory, and can be used for 3D reconstruction and relighting.
SVIP: Semantically Contextualized Visual Patches for Zero-Shot Learning
Zhi Chen (University of Southern Queensland), Zi Huang (University of Queensland)
ClassificationRecognitionTransformerImage
🎯 What it does: This paper proposes a zero-shot learning framework based on Vision Transformer, called SVIP, which addresses the semantic mismatch problem by identifying and processing semantically irrelevant image patches during the input stage.
SVTRv2: CTC Beats Encoder-Decoder Models in Scene Text Recognition
Yongkun Du (Fudan University), Yu-Gang Jiang (Fudan University)
RecognitionTransformerText
🎯 What it does: This paper presents SVTRv2, an improved scene text recognition model with a CTC structure that significantly enhances the handling of irregular text and language context while maintaining fast inference characteristics.
SweetTok: Semantic-Aware Spatial-Temporal Tokenizer for Compact Video Discretization
Zhentao Tan (Kuaishou Technology), Peng Jiang (Kuaishou Technology)
RecognitionGenerationCompressionGraph Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkImageVideo
🎯 What it does: We propose SweetTok, a tokenizer that decouples spatial and temporal information in videos, compresses it through a query autoencoder, and achieves high compression rates and high-fidelity reconstruction and generation by combining a motion-enhanced language dictionary.
Switch-a-View: View Selection Learned from Unlabeled In-the-wild Videos
Sagnik Majumder (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)
ClassificationRecognitionRetrievalDomain AdaptationTransformerLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes SWITCH-A-VIEW, a method for learning human perspective switching patterns from a vast amount of unlabeled real-world 'how-to' videos and transferring these patterns to the automatic perspective selection task in multi-view videos.
SynAD: Enhancing Real-World End-to-End Autonomous Driving Models through Synthetic Data Integration
Jongsuk Kim (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
Data SynthesisAutonomous DrivingTransformerDiffusion modelPoint Cloud
🎯 What it does: Construct synthetic traffic scenarios and integrate them into a real end-to-end autonomous driving model to enhance its robustness in diverse driving situations.
SyncDiff: Synchronized Motion Diffusion for Multi-Body Human-Object Interaction Synthesis
Wenkun He (Tsinghua University), Li Yi (Tsinghua University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: The SyncDiff framework is proposed for generating multi-body human-object interaction actions, capable of synchronizing the motion trajectories of multiple people, hands, and rigid bodies.
Synchronization of Multiple Videos
Avihai Naaman (Ben Gurion University of the Negev), Oren Freifeld (Ben Gurion University of the Negev)
RetrievalRepresentation LearningAuto EncoderVideo
🎯 What it does: A new Temporal Prototype Learning (TPL) framework is proposed for global synchronization of multiple different scenes or AI-generated videos without reference videos.
Synchronizing Task Behavior: Aligning Multiple Tasks during Test-Time Training
Wooseong Jeong (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
SegmentationDepth EstimationDomain AdaptationTransformerAuto EncoderImage
🎯 What it does: To address the domain shift problem in multi-task learning, a novel test-time training method called S4T is proposed, which can synchronize the adaptation process of different tasks in the target domain, enhancing overall performance.
SynCity: Training-Free Generation of 3D Worlds
Paul Engstler (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: This paper presents SynCity, a method for generating freely navigable three-dimensional worlds using pre-trained language models, two-dimensional image generators, and three-dimensional generators, achieved through unsupervised prompt engineering.
Synergistic Prompting for Robust Visual Recognition with Missing Modalities
Zhihui Zhang (Beijing Institute of Technology), Xiaoshuai Hao (Beijing Academy of Artificial Intelligence)
RecognitionTransformerPrompt EngineeringMultimodality
🎯 What it does: This paper proposes the Synergistic Prompting (SyP) framework to address the performance degradation caused by missing modalities in multimodal visual recognition tasks.
SynFER: Towards Boosting Facial Expression Recognition with Synthetic Data
Xilin He (Shenzhen University), Xiangyu Yue (Shenzhen University)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: A synthetic data generation framework named SynFER based on diffusion models has been constructed to generate facial expression images with reliable emotional labels.
SynTag: Enhancing the Geometric Robustness of Inversion-based Generative Image Watermarking
Han Fang (National University of Singapore), Ee-Chien Chang (National University of Singapore)
RestorationGenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: A synchronous tag injection method (SynTag) is proposed, which embeds template features sensitive to geometric distortion in an inversion-based generative model, and then uses this template to restore image geometric distortion and extract watermarks.
Synthesizing Near-Boundary OOD Samples for Out-of-Distribution Detection
Jinglun Li (Fudan University), Wenqiang Zhang (Fudan University)
Data SynthesisAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringDiffusion modelImageBenchmark
🎯 What it does: By using a multimodal large language model to extract contextual labels, we iteratively generate near-boundary OOD samples in the image feature space using a diffusion model, and fine-tune the image encoder and text negative label features of CLIP with these synthetic samples, thereby improving OOD detection performance.
Synthetic Video Enhances Physical Fidelity in Video Synthesis
Qi Zhao, Bohan Wang
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideo
🎯 What it does: This paper studies how to enhance the physical realism of video generation models using synthetic videos rendered with Blender and Unreal Engine, and validates this approach on three types of tasks: large human motion, wide-angle camera rotation, and layer separation.
T2Bs: Text-to-Character Blendshapes via Video Generation
Jiahao Luo (University of California), Jian Wang (Snap Inc.)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: An end-to-end framework for animatable 3D character head blend shapes from text (T2Bs) is proposed, combining text-to-3D static generation, 4D video diffusion, and deformable 3D Gaussian scattering to achieve high-quality, controllable facial animations.
T2I-Copilot: A Training-Free Multi-Agent Text-to-Image System for Enhanced Prompt Interpretation and Interactive Generation
Chieh-Yun Chen (SHI Labs Georgia Tech), Humphrey Shi (SHI Labs Georgia Tech)
GenerationData SynthesisTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextMultimodalityBenchmark
🎯 What it does: A training-free multi-agent system is proposed, which automatically parses prompts, selects models, and iteratively optimizes the text-to-image generation process.
TAB: Transformer Attention Bottlenecks enable User Intervention and Debugging in Vision-Language Models
Pooyan Rahmanzadehgervi (Auburn University), Anh Totti Nguyen (Auburn University)
Object DetectionExplainability and InterpretabilityTransformerVision Language ModelImageText
🎯 What it does: A single-head Transformer Attention Bottleneck (TAB) layer was designed and implemented, inserted into a vision-language model for image difference description, enabling the model to provide both difference explanations and interpretable, editable attention.
TACO: Taming Diffusion for in-the-wild Video Amodal Completion
Ruijie Lu (Peking University), Siyuan Huang (BIGAI)
RestorationGenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: Proposes the TACO model to achieve coherent and visible reconstruction of occluded objects in videos.
TAD-E2E: A Large-scale End-to-end Autonomous Driving Dataset
Chang Liu (Tencent), Kuifeng Su (Tencent)
Autonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: An end-to-end autonomous driving dataset for large-scale and complex urban scenarios, TAD-E2E, is proposed. Existing state-of-the-art (SOTA) end-to-end methods are reproduced and evaluated on this dataset, followed by the design and validation of a multi-modal sparse end-to-end network, SparseFusion-E2E (SPF-E2E), as a baseline.