ICCV 2025 Papers — Page 21
IEEE/CVF International Conference on Computer Vision · 2701 papers
RnGCam: High-speed video from rolling & global shutter measurements
Kevin Tandi (University of California), Nick Antipa (University of California)
RestorationGenerationCompressionOptical FlowVideo
🎯 What it does: This paper proposes a dual-camera system RnGCam, which utilizes low-cost rolling shutter cameras combined with a scatterer and global shutter camera optical encoding to fuse two measurements for reconstructing high-resolution video at thousands of frames per second.
ROADWork: A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones
Anurag Ghosh (Carnegie Mellon University), Srinivasa G. Narasimhan (Carnegie Mellon University)
Object DetectionSegmentationAutonomous DrivingVision Language ModelImageVideoBenchmark
🎯 What it does: A ROADWork dataset and benchmark are proposed and released for training models to recognize, observe, analyze, and safely drive through construction zones.
ROAR: Reducing Inversion Error in Generative Image Watermarking
Hanyi Wang (Shanghai Jiao Tong University), Ee-Chien Chang (National University of Singapore)
GenerationOptimizationMixture of ExpertsDiffusion modelImage
🎯 What it does: This study proposes the ROAR framework, which reduces errors in the reverse process through dual-domain optimization, significantly enhancing the robustness of generated image watermarks.
RobAVA: A Large-scale Dataset and Baseline Towards Video based Robotic Arm Action Understanding
Baoli Sun (Dalian University of Technology), Zhiyong Wang (Dalian University of Technology)
RecognitionAnomaly DetectionRobotic IntelligenceTransformerLarge Language ModelContrastive LearningVideo
🎯 What it does: A large-scale robot arm action video dataset called RobAVA is proposed, along with a motion recognition model based on atomic attributes, AGPT-Net.
Robin3D: Improving 3D Large Language Model via Robust Instruction Tuning
Weitai Kang (University of Illinois Chicago), Yan Yan (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Proposes the Robin3D 3D large language model and trains it with 1 million instruction pairs generated by RIG;
RoboAnnotatorX: A Comprehensive and Universal Annotation Framework for Accurate Understanding of Long-horizon Robot Demonstration
Longxin Kou (Tianjin University), Jianye Hao (Tianjin University)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: This paper presents RoboAnnotatorX, which utilizes a multi-scale token-efficient encoder and MLLM to automatically generate fine-grained, temporally consistent annotations for long-duration robot demonstrations.
RoboFactory: Exploring Embodied Agent Collaboration with Compositional Constraints
Yiran Qin (Sun Yat-sen University), Lei Bai (Shanghai Jiao Tong University)
Robotic IntelligenceLarge Language ModelReinforcement LearningDiffusion modelMultimodalityBenchmark
🎯 What it does: A framework of 'combinatorial constraints' (logical, spatial, temporal) and an automated data collection platform called RoboFactory are proposed for multi-body robotic collaborative tasks, along with the creation of the first multi-body manipulation benchmark based on ManiSkill.
RoboPearls: Editable Video Simulation for Robot Manipulation
Tang Tao (Shenzhen Campus of Sun Yat-sen University), Xiaodan Liang (Shenzhen Campus of Sun Yat-sen University)
Robotic IntelligenceLarge Language ModelReinforcement LearningVision Language ModelGaussian SplattingVideo
🎯 What it does: A video simulation framework named RoboPearls is proposed, which can construct high-quality dynamic scenes based on 3D Gaussian Splatting from demonstration videos. It supports automatic generation and editing of simulation scenes through natural language interaction to enhance the effectiveness of robot manipulation learning.
RoboTron-Drive: All-in-One Large Multimodal Model for Autonomous Driving
Zhijian Huang, Lin Ma
Autonomous DrivingTransformerLarge Language ModelPrompt EngineeringImageVideoMultimodalityPoint Cloud
🎯 What it does: A full-scene multimodal autonomous driving model, RoboTron-Drive, is proposed, capable of receiving various sensor inputs (single-view/multi-view images, videos, LiDAR) and completing 13 tasks including perception, prediction, and planning.
RoboTron-Mani: All-in-One Multimodal Large Model for Robotic Manipulation
Feng Yan (Meituan), Lin Ma (Meituan)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningImageTextMultimodality
🎯 What it does: RoboTron-Mani is proposed, a multimodal large model capable of accepting multi-view images, camera parameters, and text instructions, and outputting actions, images, and occupancy maps, while also constructing a unified RoboData dataset.
RoboTron-Nav: A Unified Framework for Embodied Navigation Integrating Perception, Planning, and Prediction
Yufeng Zhong (Meituan), Lin Ma
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: A unified framework called RoboTron-Nav is proposed, integrating perception, planning, and prediction, and achieving long-distance efficient navigation through multi-task collaboration (navigation + entity question answering) and adaptive 3D perception historical sampling.
RoboTron-Sim: Improving Real-World Driving via Simulated Hard-Case
Baihui Xiao (Meituan), Lin Ma (Meituan)
Autonomous DrivingLarge Language ModelPrompt EngineeringAuto EncoderImageMultimodality
🎯 What it does: Utilizing simulated hard cases to enhance real-world driving performance;
RoBridge: A Hierarchical Architecture Bridging Cognition and Execution for General Robotic Manipulation
Kaidong Zhang (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper proposes RoBridge, a method that bridges the descriptive capabilities of large-scale visual language models with the procedural capabilities of reinforcement learning through a hierarchical architecture, enabling general robotic manipulation.
Robust 3D Object Detection using Probabilistic Point Clouds from Single-Photon LiDARs
Bhavya Goyal (University of Wisconsin), Mohit Gupta (University of Wisconsin)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Using the raw temporal histogram of single-photon LiDAR, a probability point cloud (PPC) with confidence (probability) attributes is generated, and based on this, a lightweight method (NPD filtering, FPPS) is designed to achieve robust 3D object detection without modifying the network structure.
Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling
Hayeon Kim (Seoul National University), Se Young Chun (Seoul National University)
SegmentationGenerationScore-based ModelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes RoMaP, which utilizes geometry-aware soft label segmentation and regularized Score Distillation Sampling to achieve precise part-level editing of 3D Gaussian scenes.
Robust Adverse Weather Removal via Spectral-based Spatial Grouping
Yuhwan Jeong (KAIST), Kuk-Jin Yoon (KAIST)
RestorationTransformerImage
🎯 What it does: This paper proposes SSGformer, a unified multi-weather image denoising model that simultaneously restores various adverse weather conditions such as rain, snow, fog, and raindrops through spectral decomposition and spatial grouping attention.
Robust and Efficient 3D Gaussian Splatting for Urban Scene Reconstruction
Zhensheng Yuan (Jinan University), Guanghua Yang (Jinan University)
GenerationComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: A robust and efficient 3D Gaussian splatting reconstruction and real-time rendering framework aimed at urban-scale scenarios is proposed.
Robust Dataset Condensation using Supervised Contrastive Learning
Nicole Hee-Yeon Kim (Korea Advanced Institute of Science and Technology), Hwanjun Song (Korea Advanced Institute of Science and Technology)
Data SynthesisCompressionContrastive LearningImage
🎯 What it does: A robust dataset compression method called RDC is proposed, which can still generate high-quality synthetic data under noisy labels.
Robust Low-light Scene Restoration via Illumination Transition
Ze Li (Hong Kong University of Science and Technology), P. Y. Mok (Hong Kong University of Science and Technology)
RestorationNeural Radiance FieldImage
🎯 What it does: Proposes the RoSe framework, which utilizes low-light multi-view images to achieve reconstruction of normally lit scenes and new view rendering through 3D illumination transfer estimation;
Robust Machine Unlearning for Quantized Neural Networks via Adaptive Gradient Reweighting with Similar Labels
Yujia Tong (Wuhan University of Technology), Chuang Hu (University of Macau)
OptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A machine forgetting framework called QMUL specifically designed for low-bit-width quantized networks is proposed, addressing the issues of noise amplification and gradient imbalance that occur during the forgetting process of quantized models.
Robust Multi-View Learning via Representation Fusion of Sample-Level Attention and Alignment of Simulated Perturbation
Jie Xu (University of Electronic Science and Technology of China), Xiaofeng Zhu (Southeast University)
ClassificationRetrievalRepresentation LearningTransformerContrastive LearningMultimodality
🎯 What it does: A robust multi-view learning framework RML is proposed, which utilizes sample-level attention to fuse multi-view representations and achieves representation alignment and robustness enhancement through contrastive learning with simulated noise and unavailable perturbations.
Robust Test-Time Adaptation for Single Image Denoising Using Deep Gaussian Prior
Qing Ma (Hong Kong Polytechnic University), Zhe Peng (Hong Kong Polytechnic University)
RestorationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes an adaptive framework for single image denoising based on deep Gaussian priors, where a Gaussian denoising model is first pre-trained using a self-supervised approach. Then, for each image to be processed, a pixel pool is constructed, and pseudo-instances are randomly sampled for a small number of iterative fine-tuning, allowing the model to quickly adapt to unknown noise distributions.
Robust Unfolding Network for HDR Imaging with Modulo Cameras
Zhile Chen (South China University of Technology), Hui Ji (National University of Singapore)
RestorationSpiking Neural NetworkImage
🎯 What it does: Using low dynamic range images captured by a modulated camera, high dynamic range images are reconstructed through the inversion of modulation operations.
RobuSTereo: Robust Zero-Shot Stereo Matching under Adverse Weather
Yuran Wang (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
Data SynthesisDepth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: The RobuSTereo framework is proposed, which improves zero-shot stereo matching performance in adverse weather conditions such as rain and fog by augmenting adversarial weather data and enhancing the feature encoder.
Robustifying Zero-Shot Vision Language Models by Subspaces Alignment
Junhao Dong (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
RetrievalAdversarial AttackTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This study improves the zero-shot adversarial robustness of VLM by aligning the subspaces generated from image enhancement sets and synonym/antonym text prompts, achieving adversarial fine-tuning from sample level to distribution level.
RobustSplat: Decoupling Densification and Dynamics for Transient-Free 3DGS
Chuanyu Fu (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
RestorationData SynthesisOptimizationNeural Radiance FieldGaussian SplattingPoint CloudBenchmark
🎯 What it does: This paper proposes RobustSplat, a robust optimization solution to address the artifacts produced by 3D Gaussian Splatting in scenes with transient objects.
RoCo-Sim: Enhancing Roadside Collaborative Perception through Foreground Simulation
Yuwen Du (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
Data SynthesisAutonomous DrivingOptimizationImage
🎯 What it does: Developed the RoCo-Sim simulation framework for generating multi-view consistent foreground insertions and realistic training data from a single fixed-view image.
RogSplat: Robust Gaussian Splatting via Generative Priors
Hanyang Kong (National University of Singapore), Xinchao Wang (National University of Singapore)
RestorationOptimizationDiffusion modelRectified FlowNeural Radiance FieldGaussian SplattingImage
🎯 What it does: In uncontrolled real-world images with occlusions and dynamic objects, high-quality 3D scenes are reconstructed using 3D Gaussian splatting (3DGS), while simultaneously identifying and repairing occluded areas during the optimization process.
RomanTex: Decoupling 3D-aware Rotary Positional Embedded Multi-Attention Network for Texture Synthesis
Yifei Feng (Tencent), Chunchao Guo (Tencent)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Based on a multi-view diffusion model combined with 3D geometric information, high-quality, cross-view consistent 3D textures are generated.
RoMo: Robust Motion Segmentation Improves Structure from Motion
Lily Goli (Google DeepMind), Andrea Tagliasacchi (University of Toronto)
SegmentationPose EstimationOptical FlowVideo
🎯 What it does: This paper proposes a zero-shot video motion segmentation method called RoMo, which iteratively improves the motion mask by combining optical flow and disparity geometric constraints with features from a pre-trained video segmentation model, enhancing the structure from motion (SfM) camera pose estimation for dynamic scenes.
Ross3D: Reconstructive Visual Instruction Tuning with 3D-Awareness
Haochen Wang (National Laboratory of Pattern Recognition, Multimedia and Artificial Intelligence School, Chinese Academy of Sciences), Zhaoxiang Zhang (National Laboratory of Pattern Recognition, Multimedia and Artificial Intelligence School, Chinese Academy of Sciences)
RecognitionGenerationTransformerLarge Language ModelDiffusion modelMultimodalityPoint Cloud
🎯 What it does: ROSS3D is proposed, a general method to enhance the understanding of indoor 3D scenes by large multimodal models (LMM) through a 3D perception visual pre-training task.
ROVI: A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation
Cihang Peng (State Key Lab of CAD and CG Zhejiang University), Kun Zhou (State Key Lab of CAD and CG Zhejiang University)
Object DetectionGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
🎯 What it does: A text-to-image generation dataset for instance localization, ROVI, has been constructed, and the GLIGEN model has been trained based on it.
RS-vHeat: Heat Conduction Guided Efficient Remote Sensing Foundation Model
Huiyang Hu (Aerospace Information Research Institute, Chinese Academy of Sciences), Xian Sun (Aerospace Information Research Institute, Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerContrastive LearningImageMultimodality
🎯 What it does: This paper proposes RS-vHeat, a multimodal remote sensing foundational model based on the heat diffusion equation, which simulates thermal flow in remote sensing images using a heat conduction operator, balancing global receptive fields with efficient computation.
RTMap: Real-Time Recursive Mapping with Change Detection and Localization
Yuheng Du (CaiNiao Inc.), Qiang Li (CaiNiao Inc.)
Autonomous DrivingOptimizationTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: RTMap is proposed, an end-to-end real-time online high-definition map construction framework that integrates map localization, structural change detection, and multi-vehicle collaborative updates.
S2M2: Scalable Stereo Matching Model for Reliable Depth Estimation
Junhong Min (Samsung Electronics), Minyong Choi (Samsung Electronics)
Depth EstimationTransformerImage
🎯 What it does: A scalable global stereo matching model S²M² is proposed, which employs multi-resolution Transformer, Adaptive Gated Fusion, and optimal transport-based global matching with Probability Mode Concentration (PMC) loss to achieve reliable depth estimation for high-resolution, multi-disparity range scenes.
S3E: Self-Supervised State Estimation for Radar-Inertial System
Shengpeng Wang (Huazhong University of Science and Technology), Wei Wang (Wuhan University)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A self-supervised S3E radar-inertial system state estimation method is proposed, which fuses the amplitude spectrum of millimeter-wave radar with IMU data to directly extract geometrically consistent feature points from the radar spectrum and estimate the vehicle's instantaneous speed.
S3R-GS: Streamlining the Pipeline for Large-Scale Street Scene Reconstruction
Guangting Zheng (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
Autonomous DrivingComputational EfficiencyGaussian SplattingPoint CloudOrdinary Differential Equation
🎯 What it does: A street view reconstruction framework S3R-GS based on 3D Gaussian Splatting is proposed, significantly reducing the computational load and time for large-scale street view reconstruction.
S4M: Boosting Semi-Supervised Instance Segmentation with SAM
Heeji Yoon (KAIST), Seungryong Kim (KAIST)
Object DetectionSegmentationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: By transferring the localization capability of the Segment Anything Model (SAM) to a semi-supervised instance segmentation framework, and enhancing model performance through structured distillation, pseudo-label refinement, and augmentation techniques.
SA-LUT: Spatial Adaptive 4D Look-Up Table for Photorealistic Style Transfer
Zerui Gong (Nanyang Technological University), Chen Change Loy (SenseTime Research)
Image TranslationGenerationGenerative Adversarial NetworkImageVideoBenchmark
🎯 What it does: This work proposes a spatially adaptive 4D lookup table (SA-LUT) for achieving high-quality, real-time style transfer with realistic lighting.
SA-Occ: Satellite-Assisted 3D Occupancy Prediction in Real World
Chen Chen (Chinese Academy of Sciences), Xian Sun (Chinese Academy of Sciences)
SegmentationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: The SA-Occ model is proposed, which integrates satellite imagery with road surface perspectives to achieve 3D occupancy prediction.
SAC-GNC: SAmple Consensus for adaptive Graduated Non-Convexity
Valter Piedade (Instituto Superior Tecnico), Pedro Miraldo (Mitsubishi Electric Research Labs)
Pose EstimationOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: An adaptive stepwise annealing method SAC-GNC based on sample consistency is proposed to improve the convergence and efficiency of traditional GNC in robust estimation.
Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks
Jiawei Wang (University of Science and Technology of China), Kin-Man Lam (Hong Kong Polytechnic University)
Adversarial AttackTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes the Robust-VLGuard training scheme by constructing the Robust-VLGuard visual-text security dataset and performing noise augmentation with LoRA fine-tuning. It then designs the DiffPure-VLM defense pipeline by combining the noise distribution transformation of diffusion models, both of which enhance the robustness and security of VLM against Gaussian noise and optimized visual perturbations.
SAFER: Sharpness Aware layer-selective Finetuning for Enhanced Robustness in vision transformers
Bhavna Gopal (Duke University), Yiran Chen (Duke University)
ClassificationAdversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: SAFER is proposed—a layer selective fine-tuning method for visual Transformers, aimed at alleviating adversarial overfitting and improving the accuracy of clear images and adversarial samples.
SAFT: Shape and Appearance of Fabrics from Template via Differentiable Physical Simulations from Monocular Video
David Stotko (University of Bonn), Reinhard Klein (University of Bonn)
OptimizationVideoMeshOrdinary Differential Equation
🎯 What it does: Using monocular RGB video combined with shape templates, physical simulation, and differentiable rendering, we achieve 3D dynamic reconstruction and material estimation of fabrics.
SAGI: Semantically Aligned and Uncertainty Guided AI Image Inpainting
Paschalis Giakoumoglou (Aristotle University of Thessaloniki), Panagiotis C. Petrantonakis (Aristotle University of Thessaloniki)
RestorationGenerationLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImage
🎯 What it does: A model-agnostic SAGI pipeline is proposed, which automatically generates semantically aligned prompts using LLM and conducts authenticity assessment guided by uncertainty with VLM, generating high-quality images that are difficult for humans to distinguish, and based on this, a large-scale SAGI-D dataset is constructed.
SALAD -- Semantics-Aware Logical Anomaly Detection
Matic Fučka (University of Ljubljana), Danijel Skočaj (University of Ljubljana)
Anomaly DetectionImage
🎯 What it does: A logic anomaly detection framework named SALAD is proposed, which incorporates a composition branch that explicitly models the distribution of object compositions and achieves the discrimination of logical anomalies through synthetic anomaly training.
Saliency-Aware Quantized Imitation Learning for Efficient Robotic Control
Seongmin Park (Hanyang University), Jungwook Choi (Hanyang University)
Autonomous DrivingComputational EfficiencyKnowledge DistillationRobotic IntelligenceReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: This paper proposes a low-precision quantization simulation learning framework SQIL based on attention weights, which can significantly reduce computational power and energy consumption while maintaining decision accuracy.
Salvaging the Overlooked: Leveraging Class-Aware Contrastive Learning for Multi-Class Anomaly Detection
Lei Fan (University of New South Wales), Yang Song (University of New South Wales)
Anomaly DetectionContrastive LearningImage
🎯 What it does: This paper improves the multi-class generalization performance of reconstruction models in multi-class anomaly detection by introducing category-aware contrastive learning.
SAM Encoder Breach by Adversarial Simplicial Complex Triggers Downstream Model Failures
Yi Qin (Beijing Jiaotong University), Liping Jing (Beijing Jiaotong University)
SegmentationDomain AdaptationAdversarial AttackImage
🎯 What it does: A transferable adversarial attack method called VeSCA is proposed, which utilizes the transferability of the SAM encoder to generate adversarial samples by constructing an Adversarial Simplicial Complex, thereby assessing the security of SAM in various downstream tasks.
SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree
Shuangrui Ding (Chinese University of Hong Kong), Jiaqi Wang (Shanghai AI Laboratory)
Object TrackingSegmentationVideoBenchmark
🎯 What it does: This paper proposes SAM2Long, a training-independent improvement scheme that enhances the robustness and accuracy of SAM2 in long video segmentation through tree-shaped memory and uncertainty handling.
SAM4D: Segment Anything in Camera and LiDAR Streams
Jianyun Xu (CaiNiao Inc), Qiang Li (Zhejiang University)
SegmentationAutonomous DrivingTransformerPrompt EngineeringVideoMultimodalityPoint Cloud
🎯 What it does: We propose SAM4D, a foundational model for cross-modal four-dimensional segmentation of camera and LiDAR streams that can be prompted, and construct a large-scale Waymo-4DSeg dataset.
SAME: Learning Generic Language-Guided Visual Navigation with State-Adaptive Mixture of Experts
Gengze Zhou (University of Adelaide), Qi Wu (University of Adelaide)
Robotic IntelligenceTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: A general language-guided visual navigation model called SAME is proposed, which is based on state-adaptive expert mixture and can complete seven different granularity navigation tasks within a single framework.
SAMO: A Lightweight Sharpness-Aware Approach for Multi-Task Optimization with Joint Global-Local Perturbation
Hao Ban (University at Buffalo), Kaiyi Ji (University at Buffalo)
OptimizationImage
🎯 What it does: A lightweight SAMO method is proposed, utilizing joint global-local perturbations to improve the optimization process of multi-task learning (MTL).
SAMora: Enhancing SAM through Hierarchical Self-Supervised Pre-Training for Medical Images
Shuhang Chen (Zhejiang University), Dong Ni (Zhejiang University)
SegmentationKnowledge DistillationTransformerContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: For medical image segmentation, the Segment Anything Model (SAM) is fine-tuned without prompts through multi-level self-supervised pre-training and low-rank adaptation (LoRA), and a hierarchical attention fusion module (HL-Attn) is introduced to enhance segmentation accuracy.
SAMPLE: Semantic Alignment through Temporal-Adaptive Multimodal Prompt Learning for Event-Based Open-Vocabulary Action Recognition
Jing Wang (Peking University), Tiejun Huang (Peking University)
RecognitionTransformerPrompt EngineeringContrastive LearningVideoTextMultimodality
🎯 What it does: A lightweight prompt learning framework called SAMPLE is proposed to transfer the pre-trained CLIP model to the open vocabulary action recognition task of event cameras.
SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation
Junsong Chen (NVIDIA), Enze Xie (Independent Researcher)
GenerationData SynthesisKnowledge DistillationDiffusion modelFlow-based ModelImageText
🎯 What it does: A SANA-Sprint model has been designed and implemented, capable of generating high-quality text-to-image outputs in 1-4 steps, combining continuous time consistency distillation with latent adversarial distillation.
SAS: Segment Any 3D Scene with Integrated 2D Priors
Zhuoyuan Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationKnowledge DistillationConvolutional Neural NetworkDiffusion modelPoint Cloud
🎯 What it does: By transferring knowledge from various 2D open vocabulary models to 3D point clouds, the SAS method is proposed to achieve open vocabulary segmentation for arbitrary 3D scenes.
Sat2City: 3D City Generation from A Single Satellite Image with Cascaded Latent Diffusion
Tongyan Hua (Hong Kong University of Science and Technology), Wufan Zhao (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: This paper presents Sat2City, an end-to-end framework that utilizes sparse voxel grids and hierarchical latent diffusion models to quickly generate high-quality 3D city models based solely on a single satellite image.
SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoders
Jiahui Geng (Mohamed bin Zayed University of Artificial Intelligence), Qing Li (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationExplainability and InterpretabilityTransformerVision Language ModelAuto EncoderImage
🎯 What it does: This paper proposes the SAUCE method, which utilizes sparse autoencoders to perform fine-grained erasure of specific concepts in the visual module of VLM, and achieves concept removal during inference by adjusting the corresponding features.
SC-Captioner: Improving Image Captioning with Self-Correction by Reinforcement Learning
Lin Zhang (Fudan University), Tao Chen (Fudan University)
GenerationTransformerReinforcement LearningVision Language ModelImageText
🎯 What it does: This paper proposes SC-Captioner, a multi-round self-correction framework based on reinforcement learning, which enables large vision-language models to automatically eliminate false information and supplement missing details after generating image descriptions.
SC-Lane: Slope-aware and Consistent Road Height Estimation Framework for 3D Lane Detection
Chaesong Park (Seoul National University), Jongwoo Lim (Seoul National University)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: A slope-aware and temporally consistent height map estimation framework (SC-Lane) is proposed for 3D lane detection from monocular images.
Scalable Dual Fingerprinting for Hierarchical Attribution of Text-to-Image Models
Jianwei Fei (University of Macau), Zhihua Xia (Jinan University)
GenerationData SynthesisImageText
🎯 What it does: A scalable dual fingerprint embedding method has been developed for hierarchical attribution in text-to-image models, achieving traceability for both service providers and consumers.
Scalable Image Tokenization with Index Backpropagation Quantization
Fengyuan Shi (Nanjing University), Limin Wang (Nanjing University)
RestorationGenerationData SynthesisTransformerGenerative Adversarial NetworkImage
🎯 What it does: A new scalable visual quantizer method is proposed—Index Backpropagation Quantization (IBQ), which avoids codebook collapse by globally updating all codebooks, thus enabling the training of large-scale codebooks and high-quality image reconstruction and generation.
Scalable Ranked Preference Optimization for Text-to-Image Generation
Shyamgopal Karthik (Tübingen AI Center), Anil Kag (Snap Inc.)
GenerationOptimizationReinforcement LearningDiffusion modelImageText
🎯 What it does: This paper proposes a scalable text-to-image model alignment method called RankDPO, which utilizes a completely synthetic ranking preference dataset, Syn-Pic, to fine-tune pre-trained diffusion models, resulting in improvements in prompt adherence and visual quality.
Scale Your Instructions: Enhance the Instruction-Following Fidelity of Unified Image Generation Model by Self-Adaptive Attention Scaling
Chao Zhou (University of Science and Technology of China), Nenghai Yu
GenerationTransformerAuto EncoderImageMultimodality
🎯 What it does: This study investigates the issue of unified image generation models easily overlooking instructions when handling multiple sub-instructions, and proposes an adaptive attention scaling method to improve instruction following quality.
Scaling 3D Compositional Models for Robust Classification and Pose Estimation
Xiaoding Yuan (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
ClassificationPose EstimationConvolutional Neural NetworkTransformerContrastive LearningImagePoint Cloud
🎯 What it does: This paper proposes a scalable 3D compositional network, 3D-CompNets, which significantly enhances robustness and efficiency under large-scale categories by jointly training classification and 3D pose estimation through grouped neural vertex contrastive learning with dynamic weighted combinations.
Scaling Action Detection: AdaTAD++ with Transformer-Enhanced Temporal-Spatial Adaptation
Tanay Agrawal (INRIA), Francois Bremond (INRIA)
RecognitionObject DetectionTransformerVideo
🎯 What it does: The AdaTAD++ framework is proposed, which implements long video action detection through spatial and temporal decoupled adapters and a Transformer-enhanced spatiotemporal adapter.
Scaling and Taming Adversarial Training with Synthetic Data
Juntao Wu (Jinan University), Ke Wang (Jinan University)
Data SynthesisAdversarial AttackTransformerDiffusion modelImage
🎯 What it does: The paper achieves improved adversarial robustness on ImageNet without external data through large-scale synthetic data, increased model size, and an adaptive PGD strategy.
Scaling Inference-Time Search with Vision Value Model for Improved Visual Comprehension
Xiyao Wang (University of Maryland), Lijuan Wang (Microsoft)
GenerationComputational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes VisVM (Vision Value Model), which guides the search during the reasoning phase of VLM by evaluating the long-term value of the current sentence and its potential subsequent sentences, thereby generating image descriptions with more visual details and fewer hallucinations.
Scaling Language-Free Visual Representation Learning
David Fan (Meta), Saining Xie (New York University)
Representation LearningTransformerContrastive LearningImageMultimodality
🎯 What it does: This paper explores the performance of visual self-supervised learning in multimodal tasks (VQA) by training visual self-supervised models (DINOv2, MAE) on the large-scale network image dataset MetaCLIP and expanding the model parameters from 1B to 7B, comparing it with CLIP.
Scaling Laws for Native Multimodal Models
Mustafa Shukor, Alaaeldin El-Nouby
GenerationData SynthesisComputational EfficiencyTransformerMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: This paper studies the scaling properties of native multimodal models (NMMs) trained from scratch and compares the performance of early fusion and late fusion architectures.
Scaling Omni-modal Pretraining with Multimodal Context: Advancing Universal Representation Learning Across Modalities
Yiyuan Zhang (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
GenerationRetrievalRepresentation LearningTransformerContrastive LearningImageVideoTextMultimodalityAudio
🎯 What it does: The MiCo framework is proposed, utilizing large-scale multimodal alignment data (images, depth, normal maps, audio, video, text) for omni-modal pre-training, aiming to construct general and transferable cross-modal representations.
Scaling Transformer-Based Novel View Synthesis with Models Token Disentanglement and Synthetic Data
Nithin Gopalakrishnan Nair (Johns Hopkins University), Jungyeon Park (Google)
GenerationData SynthesisDomain AdaptationTransformerDiffusion modelImage
🎯 What it does: A Token Disentangled Transformer is designed for an efficient novel perspective synthesis model without 3D priors, combined with synthetic data to achieve cross-domain generalization.
Scaling Tumor Segmentation: Best Lessons from Real and Synthetic Data
Qi Chen (Johns Hopkins University), Zongwei Zhou (Johns Hopkins University)
SegmentationData SynthesisConvolutional Neural NetworkDiffusion modelImageBiomedical DataComputed Tomography
🎯 What it does: This study investigates the data scale rule for tumor segmentation and constructs the largest and most comprehensively annotated abdominal tumor dataset, AbdomenAtlas 2.0. It also verifies the role of synthetic data in improving performance both within and outside the distribution by synthesizing tumors to accelerate model training.
SCAN: Bootstrapping Contrastive Pre-training for Data Efficiency
Yangyang Guo (National University of Singapore), Mohan Kankanhalli (National University of Singapore)
OptimizationData-Centric LearningTransformerContrastive LearningImage
🎯 What it does: A dynamic guided contrastive pre-training data pruning method called SCAN is proposed, which can adaptively select and remove unimportant or redundant data samples during the pre-training process.
ScanEdit: Hierarchically-Guided Functional 3D Scan Editing
Mohamed El Amine Boudjoghra (Technical University of Munich), Angela Dai (Technical University of Munich)
OptimizationTransformerLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: This paper proposes a hierarchical-guided instruction-driven 3D scanning editing method called ScanEdit, designed to rearrange and transform complex 3D scanned scenes from the real world.
Scendi Score: Prompt-Aware Diversity Evaluation via Schur Complement of CLIP Embeddings
Azim Ospanov (Chinese University of Hong Kong), Farzan Farnia (Chinese University of Hong Kong)
GenerationData SynthesisPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: A Schur complement decomposition method based on CLIP embeddings is proposed, defining the Scendi score to quantify model-driven diversity in generative models guided by text prompts.
Scene Coordinate Reconstruction Priors
Wenjing Bian (University of Oxford), Eric Brachmann (Niantic Spatial)
Pose EstimationDepth EstimationDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a maximum likelihood-based Scene Coordinate Regression (SCR) training framework, which enhances the quality of depth camera pose estimation and structure reconstruction by introducing various geometric priors.
Scene Graph Guided Generation: Enable Accurate Relations Generation in Text-to-Image Models via Textural Rectification
Guibao Shen (Hong Kong University of Science and Technology), Ying-cong Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: Developed the Scene Graph Adapter (SG-Adapter), which inserts a transformer adapter behind the CLIP text encoder of Stable Diffusion to correct contextual leakage in text embeddings using the triplet information from scene graphs through cross-attention; simultaneously constructed a high-quality MultiRels scene graph-image pairing dataset and proposed three evaluation metrics based on GPT-4V (SG-IoU, Relation-IoU, Entity-IoU) to quantify the correspondence between images and scene graphs.
SceneMI: Motion In-betweening for Modeling Human-Scene Interaction
Inwoo Hwang (Seoul National University), Chuan Guo (Snap Inc)
GenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: The SceneMI framework is proposed to achieve scene-aware motion intermediate frame generation, capable of synthesizing smooth and natural full-body movements given key frames and a 3D scene.
ScenePainter: Semantically Consistent Perpetual 3D Scene Generation with Concept Relation Alignment
Chong Xia (Tsinghua University), Yueqi Duan (Tsinghua University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes the ScenePainter framework, which uses a Scene Concept Graph to generate semantically consistent and visually diverse 3D view sequences from a single view, ensuring consistency through refinement during testing.
SceneSplat: Gaussian Splatting-based Scene Understanding with Vision-Language Pretraining
Yue Li (University of Amsterdam), Danda Pani Paudel (University of Amsterdam)
RecognitionSegmentationVision Language ModelGaussian SplattingPoint Cloud
🎯 What it does: This study investigates indoor scene understanding based on 3D Gaussian splatting and proposes the SceneSplat model.
SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models
Pingchuan Ma (CompVis), Björn Ommer (CompVis)
GenerationData SynthesisRetrievalFlow-based ModelContrastive LearningImageOrdinary Differential Equation
🎯 What it does: The SCFlow model is proposed, achieving implicit style-content decoupling through reversible fusion and separation of style and content via flow matching.
Scheduling Weight Transitions for Quantization-Aware Training
Junghyup Lee (Samsung Research), Bumsub Ham (Yonsei University)
Object DetectionOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A quantization-aware training method based on transfer rate (TR) scheduling is proposed, which can more accurately control the level changes of quantized weights.
SciVid: Cross-Domain Evaluation of Video Models in Scientific Applications
Yana Hasson (Google DeepMind), Andrew Zisserman (Google DeepMind)
ClassificationObject TrackingDomain AdaptationSupervised Fine-TuningVideoBenchmark
🎯 What it does: This paper constructs the SCIVID benchmark, covering five scientific video tasks including animal behavior classification, tissue tracking, weather and pressure prediction, and systematically evaluates the transfer effects of various video foundation models (ViFMs) on these tasks;
SCORE: Scene Context Matters in Open-Vocabulary Remote Sensing Instance Segmentation
Shiqi Huang (Nanyang Technological University), Bihan Wen (Nanyang Technological University)
Object DetectionSegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes the SCORE framework, which implements open vocabulary-based remote sensing instance segmentation and enhances segmentation accuracy through multi-granularity scene context.
ScoreHOI: Physically Plausible Reconstruction of Human-Object Interaction via Score-Guided Diffusion
Ao Li (Tsinghua University), Yansong Tang (Tsinghua University)
GenerationOptimizationDiffusion modelScore-based ModelImage
🎯 What it does: The ScoreHOI method based on diffusion models achieves 3D reconstruction of human-object interactions from a single RGB image, and improves reconstruction accuracy and physical plausibility through physical constraints and contact iterative optimization.
Scoring, Remember, and Reference: Catching Camouflaged Objects in Videos
Yu'ang Feng (Fudan University), Wenqiang Zhang (Fudan University)
Object DetectionSegmentationTransformerVideo
🎯 What it does: An end-to-end video camouflage object detection framework SRRNet is proposed, which utilizes memory reference frames, a dual-use decoder, and self-supervised score prediction to achieve segmentation and localization of camouflaged objects in videos.
Sculpting Memory: Multi-Concept Forgetting in Diffusion Models via Dynamic Mask and Concept-Aware Optimization
Gen Li (Clemson University), Xiaolong Ma (University of Arizona)
GenerationOptimizationKnowledge DistillationDiffusion modelImage
🎯 What it does: A multi-concept forgetting framework for diffusion models is proposed, which can selectively 'forget' multiple target concepts without significantly affecting the overall generation quality of the model, and after forgetting, maps them to corresponding higher-level or opposing concepts.
SD2Actor: Continuous State Decomposition via Diffusion Embeddings for Robotic Manipulation
Jiayi Li (Xi'an Jiaotong University)
Robotic IntelligenceLarge Language ModelReinforcement LearningDiffusion modelSequential
🎯 What it does: This paper proposes SD 2 Actor, a zero-shot language-conditioned robotic manipulation framework based on diffusion models, capable of accurately generating robot actions in continuous state spaces.
SDFit: 3D Object Pose and Shape by Fitting a Morphable SDF to a Single Image
Dimitrije Antić (University of Amsterdam), Dimitrios Tzionas (University of Amsterdam)
Object DetectionPose EstimationOptimizationImage
🎯 What it does: The SDFit framework is proposed, utilizing the mSDF model to simultaneously recover the 3D pose and shape of objects from a single image.
SDFormer: Vision-based 3D Semantic Scene Completion via SAM-assisted Dual-channel Voxel Transformer
Yujie Xue (Hunan University), Ruihui Li (Hunan University)
SegmentationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes SDFormer, a dual-channel voxel Transformer framework that utilizes the Segment Anything Model (SAM) prior for Semantic Scene Completion (SSC).
SDMatte: Grafting Diffusion Models for Interactive Matting
Longfei Huang (Shanghai University), Peng-Tao Jiang (vivo Mobile Communication Co)
SegmentationGenerationDiffusion modelImage
🎯 What it does: This paper proposes an interactive image matting method called SDMatte based on diffusion models, which utilizes visual prompts (points, boxes, masks) to achieve precise Alpha Matte prediction.
Seal Your Backdoor with Variational Defense
Ivan Sabolić (University of Zagreb), Siniša Šegvić (University of Zagreb)
OptimizationAdversarial AttackImage
🎯 What it does: A model-agnostic VIBE framework is proposed, treating clean labels as latent variables through variational inference, and recovering pseudo-labels during training using the EM algorithm and entropy-regularized optimal transport to resist backdoor attacks.
SEAL: Semantic Aware Image Watermarking
Kasra Arabi (New York University), Niv Cohen (New York University)
Data SynthesisAdversarial AttackDiffusion modelImage
🎯 What it does: A database-free diffusion model watermarking method based on image semantics, called SEAL, is proposed, which encodes semantic information using initial noise to embed watermarks without disrupting the image distribution.
Seam360GS: Seamless 360deg Gaussian Splatting from Real-World Omnidirectional Images
Changha Shin (Yonsei University), Seon Joo Kim (Yonsei University)
GenerationData SynthesisOptimizationGaussian SplattingImage
🎯 What it does: A panoramic rendering framework is proposed that combines fisheye camera distortion modeling with 3D Gaussian splatting to achieve seamless synthesis of 360° views.
SeaS: Few-shot Industrial Anomaly Image Generation with Separation and Sharing Fine-tuning
Zhewei Dai (Huazhong University of Science and Technology), Yu Zhou (Huazhong University of Science and Technology)
GenerationAnomaly DetectionPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes SeaS, a unified industrial anomaly image generation model that can simultaneously generate various anomalies, realistic normal products, and precise pixel-level anomaly masks.
Secure On-Device Video OOD Detection Without Backpropagation
Shawn Li (University of Southern California), Yue Zhao (University of Southern California)
Anomaly DetectionComputational EfficiencyMeta LearningVideo
🎯 What it does: A secure OOD detection framework SecDOOD based on cloud-device collaboration is proposed, allowing edge devices to complete OOD detection without backpropagation, ensuring data privacy while considering computational resource constraints.
Seeing 3D Through 2D Lenses: 3D Few-Shot Class-Incremental Learning via Cross-Modal Geometric Rectification
Tuo Xiang (South China University of Technology), Shengfeng He (Singapore Management University)
ClassificationRecognitionTransformerContrastive LearningPoint Cloud
🎯 What it does: For few-shot incremental learning of 3D point clouds, a Cross-Modal Geometric Correction (CMGR) framework is proposed to achieve hierarchical alignment of 3D features with the semantic intermediate layers of CLIP, and to further enhance geometric consistency and memory stability through texture augmentation and a base-new classifier.
Seeing and Seeing Through the Glass: Real and Synthetic Data for Multi-Layer Depth Estimation
Hongyu Wen (Princeton University), Jia Deng (Princeton University)
Depth EstimationConvolutional Neural NetworkImageBenchmark
🎯 What it does: Proposes a multi-layer depth estimation task that simultaneously predicts the multi-layer depth of the front and back surfaces of transparent objects and the objects behind them using a single RGB image.