ICCV 2025 Papers — Page 22
IEEE/CVF International Conference on Computer Vision · 2701 papers
Seeing the Trees for the Forest: Rethinking Weakly-Supervised Medical Visual Grounding
Ta Duc Huy (Australian Institute for Machine Learning University of Adelaide), Vu Minh Hieu Phan (Australian Institute for Machine Learning University of Adelaide)
Object DetectionSegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageBiomedical DataComputed Tomography
🎯 What it does: A weakly supervised medical visual localization method based on Disease-Aware Prompting (DAP) is proposed, achieving localization of disease regions in chest X-ray images without the need for pixel-level annotations.
Seeing the Unseen: A Semantic Alignment and Context-Aware Prompt Framework for Open-Vocabulary Camouflaged Object Segmentation
Peng Ren (Jilin University), Fuming Sun (Dalian Minzu University)
Object DetectionSegmentationVision Language ModelContrastive LearningImageBenchmark
🎯 What it does: Proposes the SuCLIP framework to address the semantic confusion problem in open vocabulary hidden object segmentation.
Seeing Through Deepfakes: A Human-Inspired Framework for Multi-Face Detection
Juan Hu (National University of Singapore), Terence Sim (National University of Singapore)
Object DetectionTransformerContrastive LearningVideo
🎯 What it does: This paper studies multi-faceted deepfake video detection and proposes a human cognition-based multi-faceted detection framework called HICOM.
SEGA: A Stepwise Evolution Paradigm for Content-Aware Layout Generation with Design Prior
Haoran Wang, Xinyan Xiao
GenerationTransformerLarge Language ModelVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: A content-aware layout generation framework SEGA based on hierarchical thinking is proposed, which breaks down the layout task into two steps: rough estimation and refinement iteration, and incorporates a feedback mechanism for design principles.
SegAnyPET: Universal Promptable Segmentation from Positron Emission Tomography Images
Yichi Zhang (Fudan University), Yuan Qi (Fudan University)
SegmentationTransformerPrompt EngineeringImageBiomedical DataPositron Emission Tomography
🎯 What it does: Developed a dedicated 3D Promptable segmentation foundation model SegAnyPET for PET images and collected the largest PETS-5k dataset.
SegmentDreamer: Towards High-fidelity Text-to-3D Synthesis with Segmented Consistency Trajectory Distillation
Jiahao Zhu (Sun Yat-sen University), Yi Zhou (Sun Yat-sen University)
GenerationData SynthesisKnowledge DistillationDiffusion modelScore-based ModelGaussian SplattingTextPoint CloudOrdinary Differential Equation
🎯 What it does: Proposes the SegmentDreamer framework to address the issue of conditional guidance imbalance in existing Consistency Distillation (CD) methods for text-to-3D generation, achieving high-fidelity text-to-3D synthesis through Segmented Consistency Trajectory Distillation (SCTD).
SEGS-SLAM: Structure-enhanced 3D Gaussian Splatting SLAM with Appearance Embedding
Tianci Wen (Nankai University), Yongchun Fang (Nankai University)
OptimizationRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: The SEGS-SLAM system is proposed and implemented, utilizing Structure Enhanced 3D Gaussian Splatting (SEPM), Appearance Embedding (AfME), and Frequency Pyramid Regularization (FPR) to achieve high-quality, realistic SLAM mapping in multi-camera (monocular, stereo, RGB-D) environments.
SEHDR: Single-Exposure HDR Novel View Synthesis via 3D Gaussian Bracketing
Yiyu Li (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
RestorationGenerationData SynthesisGaussian SplattingImage
🎯 What it does: A SEHDR framework is proposed, utilizing single-exposure multi-view LDR images to achieve HDR new view synthesis through 3D Gaussian splatting.
Selective Contrastive Learning for Weakly Supervised Affordance Grounding
WonJun Moon (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
Object DetectionSegmentationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a weakly supervised grasping function localization method based on selective prototypes and pixel contrast learning, utilizing first-person and third-person images to jointly learn objects and their executable parts.
Self-Calibrated Variance-Stabilizing Transformations for Real-World Image Denoising
Sébastien Herbreteau (Univ Rennes), Michael Unser (École Polytechnique Fédérale de Lausanne)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a zero-shot image denoising framework called Noise2VST, which utilizes self-supervised learning VST and a pre-trained Gaussian denoiser to achieve real-world noise removal.
Self-Calibrating Gaussian Splatting for Large Field-of-View Reconstruction
Youming Deng (Cornell University), Paul Debevec (Netflix Eyeline Studios)
RestorationOptimizationConvolutional Neural NetworkNeural Radiance FieldGaussian SplattingImage
🎯 What it does: A self-calibrating Gaussian Splatting framework is proposed, which can directly reconstruct high-quality 3D scenes from uncalibrated panoramic fisheye images and supports nearly 180° field of view.
Self-Ensembling Gaussian Splatting for Few-Shot Novel View Synthesis
Chen Zhao (EPFL), Mathieu Salzmann (EPFL)
GenerationData SynthesisGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes Self-Ensembling Gaussian Splatting (SE-GS) to address the overfitting problem of 3D Gaussian Splatting under limited views, enhancing the quality of few-sample perspective synthesis through self-ensemble.
Self-Reinforcing Prototype Evolution with Dual-Knowledge Cooperation for Semi-Supervised Lifelong Person Re-Identification
Kunlun Xu (Peking University), Jiahuan Zhou (Peking University)
RecognitionRetrievalContrastive LearningImageBenchmark
🎯 What it does: A semi-supervised lifelong person re-identification method SPRED is proposed, which enhances the utilization of unlabeled data through self-reinforcement prototype evolution and dual knowledge collaboration.
Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics
Zhirui Gao (National University of Defense Technology), Kai Xu (National University of Defense Technology)
OptimizationRepresentation LearningGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: A self-supervised PartGS mixed representation framework is proposed, combining 2D Gaussians and tetrahedra to achieve partial perception reconstruction and decomposition of 3D objects/scenes through multi-view images.
Self-Supervised Monocular 4D Scene Reconstruction for Egocentric Videos
Chengbo Yuan (Tsinghua University), Yang Gao (Tsinghua University)
Object DetectionSegmentationDepth EstimationOptical FlowVideo
🎯 What it does: This paper proposes a self-supervised learning-based method for 4D reconstruction of monocular front-view videos, named EgoMono4D, which can simultaneously predict camera intrinsic parameters, camera poses, video depth, and confidence masks, thereby generating dense point cloud sequences within a single forward inference framework.
Self-Supervised Sparse Sensor Fusion for Long Range Perception
Edoardo Palladin (Torc Robotics), Felix Heide (Princeton University)
Object DetectionDepth EstimationAutonomous DrivingComputational EfficiencySupervised Fine-TuningMultimodalityPoint Cloud
🎯 What it does: A sparse voxel self-supervised multimodal fusion framework is proposed to achieve long-distance (250 m) perception and object detection.
Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers
Yunshan Zhong (Xiamen University), Rongrong Ji (Xiamen University)
OptimizationData-Centric LearningTransformerReinforcement LearningImage
🎯 What it does: A no-data quantization method for visual Transformers, SARDFQ, is proposed to address the issues of semantic distortion and semantic insufficiency in synthetic images.
Semantic Causality-Aware Vision-Based 3D Occupancy Prediction
Dubing Chen (University of Macau), Jianbing Shen (University of Macau)
SegmentationAutonomous DrivingConvolutional Neural NetworkSupervised Fine-TuningPoint CloudBenchmark
🎯 What it does: This paper proposes an end-to-end supervised 2D→3D semantic occupancy prediction framework based on causal loss, addressing the error propagation caused by semantic confusion in the traditional Lift-Splat-Shoot method.
Semantic Discrepancy-aware Detector for Image Forgery Identification
Ziye Wang (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
Anomaly DetectionTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImage
🎯 What it does: A semantic difference-aware image forgery detection method SDD is proposed, which aligns the semantic concepts of the CLIP visual-language model with forgery features and enhances detection performance through reconstruction learning.
Semantic Equitable Clustering: A Simple and Effective Strategy for Clustering Vision Tokens
Qihang Fan (Chinese Academy of Sciences), Ran He (Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationRetrievalComputational EfficiencyTransformerVision Language ModelImageMultimodality
🎯 What it does: A single-step equitable clustering method based on global semantic similarity (Semantic Equitable Clustering, SEC) is proposed, and it is applied to construct an efficient visual Transformer (SECViT) and a visual-language connector for multimodal large language models (such as LLaVA-1.5).
Semantic versus Identity: A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification
Yuan Tian (Shanghai AI Laboratory), Guangtao Zhai (Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University)
ClassificationRestorationSegmentationData SynthesisSafty and PrivacyTransformerDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A tunable medical image de-identification (DeID) framework is proposed, which significantly enhances privacy protection while maintaining downstream task performance by first masking identity-related areas and then compensating for the lost diagnostic semantics using a Medical Foundation Model (MFM).
Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity
Sung Ju Lee (Seoul National University), Nam Ik Cho (Seoul National University)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes a Hermitian Symmetric Fourier Watermark (SFW) and a center-aware embedding strategy to achieve semantic watermark embedding and detection in latent diffusion models;
Semantic-guided Camera Ray Regression for Visual Localization
Yesheng Zhang (Shanghai Jiao Tong University), Xu Zhao (Shanghai Jiao Tong University)
Pose EstimationTransformerImage
🎯 What it does: This paper proposes a visual localization framework called DIMM, which utilizes DINO visual Transformer to extract local and global semantic features, and achieves soft fusion of multiple mappings through semantic attention, ultimately resulting in high-precision camera pose estimation.
SemGes: Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning
Lanmiao Liu (Max Planck Institute for Psycholinguistics), Zerrin Yumak (Utrecht University)
GenerationTransformerAuto EncoderTextMultimodalityAudio
🎯 What it does: A two-stage semantic-driven co-speech gesture generation framework, SemGes, is proposed. It first learns a discrete codebook of full-body and hand movements through VQ-VAE, and then utilizes a cross-modal Transformer to integrate speech, text semantics, and speaker identity to generate semantically coherent and stylistically consistent gestures.
Semi-supervised Concept Bottleneck Models
Lijie Hu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
ClassificationRecognitionImage
🎯 What it does: A semi-supervised concept bottleneck model (SSCBM) is proposed, achieving concept prediction and task classification with a small amount of concept labeling through KNN pseudo-labels and alignment loss.
Semi-supervised Deep Transfer for Regression without Domain Alignment
Mainak Biswas (Indian Institute of Science), Devarajan Sridharan (Indian Institute of Science)
Domain AdaptationAdversarial AttackBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A source-free semi-supervised deep transfer method called CRAFT is proposed, specifically targeting domain-aligned transfer for regression tasks.
Semi-ViM: Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning
Hongyang He (University of Warwick), Victor Sanchez (University of Warwick)
ClassificationTransformerImage
🎯 What it does: A semi-supervised learning framework called Semi-ViM is proposed, based on a bidirectional state space model, replacing Transformer with Vision Mamba to address the issue of label imbalance.
SemiVisBooster: Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance
Wenjin Zhang (Rutgers University), Ivan Marsic (Rutgers University)
ClassificationContrastive LearningImage
🎯 What it does: This paper proposes SemiVisBooster, which enhances semi-supervised learning by utilizing the semantic information of label names to improve fine-grained classification performance.
SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis
Xiangyue Zhang, Zhigang Tu
GenerationData SynthesisTransformerContrastive LearningVideoMultimodality
🎯 What it does: Proposes the SemTalk framework, which divides co-speech action generation into rhythmic base actions and sparse semantic actions, and achieves semantic emphasis through adaptive fusion.
Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring
Yufei Zhu (Southeast University), Wei You (Huawei Technologies)
RestorationConvolutional Neural NetworkOptical FlowImage
🎯 What it does: A single-frame event camera deblurring method is proposed, which first separates the conflicting edge information and motion information in the event stream, and then jointly recovers the blurred image through an edge enhancement module and a motion-driven scale-adaptive deblurring module.
SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions
Mengwei Xie (Alibaba Group), Xing Wei (Xi'an Jiaotong University)
Autonomous DrivingTransformerGraph
🎯 What it does: By modeling the lane graph as an incremental graph expansion sequence, automated generation of lane topology maps is achieved.
Sequential Gaussian Avatars with Hierarchical Motion Context
Wangze Xu (Shanghai Artificial Intelligence Laboratory), Xiao Sun (Shanghai Artificial Intelligence Laboratory)
GenerationData SynthesisPose EstimationGaussian SplattingVideo
🎯 What it does: This paper proposes SeqAvatar, which utilizes 3D Gaussian Splatting combined with hierarchical motion context to achieve high-quality, real-time animatable 3D human avatars.
Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection
Anja Delić (University of Zagreb), Siniša Šegvić (University of Zagreb)
Anomaly DetectionTransformerVideoSequential
🎯 What it does: This paper proposes a skeleton sequence anomaly detection method called SeeKer, which identifies abnormal behaviors by predicting the probability distribution of skeletal key points.
SEREP: Semantic Facial Expression Representation for Robust In-the-Wild Capture and Retargeting
Arthur Josi (Ecole de Technologie Superieure), Marc-André Carbonneau (Ubisoft LaForge)
GenerationRepresentation LearningConvolutional Neural NetworkImageMeshBenchmark
🎯 What it does: This paper proposes SEREP, which can capture and remap facial expressions from single-camera images at any identity, generating a 3D mesh that conforms to the target facial geometry.
Serialization based Point Cloud Oversegmentation
Chenghui Lu (Huaqiao University), Haiyan Guan (Nanjing University of Information Science and Technology)
SegmentationOptimizationConvolutional Neural NetworkGraph Neural NetworkPoint Cloud
🎯 What it does: A point cloud refinement method SPCNet based on Hilbert curve serialization is designed, utilizing serialized initial segmentation, semantic similarity matching, and cross-attention to achieve superpoint generation and iterative optimization.
SFUOD: Source-Free Unknown Object Detection
Keon-Hee Park (Kyung Hee University), Gyeong-Moon Park (Korea University)
Object DetectionDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the Source-Free Unknown Object Detection (SFUOD) scenario and designs the CollaPAUL framework to achieve informed object transfer and unknown object detection.
SG-LDM: Semantic-Guided LiDAR Generation via Latent-Aligned Diffusion
Zhengkang Xiang (University of Melbourne), Kourosh Khoshelham (University of Melbourne)
SegmentationGenerationDomain AdaptationDiffusion modelPoint Cloud
🎯 What it does: A semantic-guided LiDAR diffusion model SG-LDM is proposed to directly convert semantic segmentation maps into high-quality LiDAR point clouds.
SGAD: Semantic and Geometric-aware Descriptor for Local Feature Matching
Xiangzeng Liu (Xidian University), Miao Fan (Navinfo Europe B.V)
Pose EstimationConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: The Semantic and Geometric-aware Descriptor Network (SGAD) is proposed, achieving efficient and accurate region-to-point matching through regional descriptors.
ShadowHack: Hacking Shadows via Luminance-Color Divide and Conquer
Jin Hu (Tianjin University), Xiaojie Guo (Tianjin University)
RestorationTransformerImage
🎯 What it does: A method is proposed to decompose shadow images into brightness and color components, and to perform brightness recovery and color reconstruction using LRNet and CRNet, respectively, to achieve shadow removal.
Shape of Motion: 4D Reconstruction from a Single Video
Qianqian Wang (Google DeepMind), Angjoo Kanazawa (University of California Berkeley)
Object TrackingGenerationData SynthesisDepth EstimationNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: Reconstruct a complete four-dimensional dynamic scene from scratch using monocular video, generating continuous three-dimensional motion trajectories and renderings that can synthesize new viewpoints.
SHeaP: Self-Supervised Head Geometry Predictor Learned via 2D Gaussians
Liam Schoneveld (Woven by Toyota), Matthias Nießner (Technical University of Munich)
RecognitionGenerationGraph Neural NetworkTransformerGaussian SplattingVideo
🎯 What it does: Under a single image, the self-supervised SHeaP model predicts both the 3DMM mesh and the associated 2D Gaussian splat, achieving real-time head geometry reconstruction and animation.
SHIFT: Smoothing Hallucinations by Information Flow Tuning for Multimodal Large Language Models
Sudong Wang (Nanyang Technological University), Xiangyang Ji (Tsinghua University)
Object DetectionGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: The SHIFT method is proposed, which identifies information mutations during the inference phase of multimodal large language models by monitoring the Jensen-Shannon Divergence between adjacent layers, and smooths the mutation layers with shallow continuous information to suppress the model's hallucination outputs.
ShortFT: Diffusion Model Alignment via Shortcut-based Fine-Tuning
Xiefan Guo (Beihang University), Di Huang (Beihang University)
GenerationData SynthesisReinforcement LearningDiffusion modelImage
🎯 What it does: By constructing a trajectory-preserving few-step diffusion model shortcut (denoising shortcut), end-to-end gradient backpropagation for the diffusion model is achieved, completing the alignment training with the reward function.
ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers
Qianhao Yuan (Institute of Software, Chinese Academy of Sciences), Le Sun (Institute of Software, Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes a method called ShortV, which is untrained and freezes visual tokens to enhance the reasoning efficiency of multimodal large language models.
Shot-by-Shot: Film-Grammar-Aware Training-Free Audio Description Generation
Junyu Xie, Andrew Zisserman
GenerationTransformerLarge Language ModelVision Language ModelVideoAudio
🎯 What it does: This paper proposes a training-agnostic two-stage audio description generation framework that utilizes shot-level temporal context and film grammar (shot scale, cue structure) to guide video LLM in generating dense descriptions, which are then compressed into standardized audio descriptions using LLM.
Sibai: A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning
Melanie Götz (University of Wurzburg), Alexandra Dmitrienko (University of Wurzburg)
Anomaly DetectionFederated LearningMeta LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A meta-classifier called Sibai based on Siamese networks is proposed in a federated learning environment to detect and filter malicious local models, preventing backdoor injection.
SIC: Similarity-Based Interpretable Image Classification with Neural Networks
Tom Nuno Wolf (Technical University of Munich), Christian Wachinger (Technical University of Munich)
ClassificationExplainability and InterpretabilityImage
🎯 What it does: A fully intrinsic interpretable neural network named SIC is proposed, which implements image classification using support vectors and similarity, and generates pixel-level explanations through B-Cos transformation.
SIGMAN: Scaling 3D Human Gaussian Generation with Millions of Assets
Yuhang Yang (University of Science and Technology of China), Junting Dong (University of Science and Technology of China)
GenerationData SynthesisPose EstimationDiffusion modelAuto EncoderGaussian SplattingImagePoint Cloud
🎯 What it does: An end-to-end 3D human high-quality Gaussian point cloud generation framework based on UV-structured VAE and MM-DiT is proposed, and a unified dataset of over one million samples, HGS-1M, is constructed.
SignRep: Enhancing Self-Supervised Sign Representations
Ryan Wong (University of Surrey), Richard Bowden (University of Surrey)
RecognitionRepresentation LearningTransformerAuto EncoderVideo
🎯 What it does: A self-supervised masked autoencoder framework called SignRep is proposed to learn representations of unlabeled sign language videos.
Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator
Ronglai Zuo (Imperial College London), Stefanos Zafeiriou (Imperial College London)
GenerationPose EstimationRetrievalTransformerLarge Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: A multilingual sign language generation model SOKE is proposed, which utilizes a pre-trained language model and discretized sign language tokens for automatic text-to-sign language generation.
SILO: Solving Inverse Problems with Latent Operators
Ron Raphaeli (Technion Israel Institute of Technology), Michael Elad (Technion Israel Institute of Technology)
RestorationGenerationDiffusion modelAuto EncoderImage
🎯 What it does: The SILO method is proposed for solving inverse problems, operating entirely within the latent space, thus avoiding the cost of switching between pixel space and latent space.
Sim-DETR: Unlock DETR for Temporal Sentence Grounding
Jiajin Tang (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
RecognitionObject DetectionTransformerVideoText
🎯 What it does: This paper proposes Sim-DETR, which addresses the query conflict issue in the temporal sentence localization task of DETR by designing two lightweight improvement schemes to construct an efficient and stable baseline model.
SiM3D: Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark
Alex Costanzino (University of Bologna), Luigi Di Stefano (University of Bologna)
SegmentationDomain AdaptationAnomaly DetectionImageMultimodalityPoint CloudBenchmark
🎯 What it does: The SiM3D benchmark is proposed to study multi-view and multi-modal 3D anomaly detection and segmentation tasks, focusing on single-instance training and the transfer from synthetic to real domains.
Similarity Memory Prior is All You Need for Medical Image Segmentation
Hao Tang (Xinjiang University), Chao Liu (Xinjiang University)
SegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A medical image segmentation network called Sim-MPNet based on similarity memory priors is proposed, along with two core modules: Dynamic Memory Weights-Loss Attention (DMW-LA) and Double-Similarity Global Internal Enhancement Module (DS-GIM).
SimMLM: A Simple Framework for Multi-modal Learning with Missing Modality
Sijie Li (University of Sheffield), Jungong Han (University of Sheffield)
ClassificationSegmentationConvolutional Neural NetworkTransformerMixture of ExpertsMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The SimMLM framework is proposed to address the issue of missing modalities during testing in multimodal learning.
SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models
Xianfu Cheng (Baidu Inc.), Zhoujun Li (Beihang University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes and releases SimpleVQA—a bilingual (English and Chinese) visual question answering benchmark designed to evaluate the factual reasoning capabilities of multimodal large language models (MLLMs). It systematically assesses 18 mainstream MLLMs and 8 text-only LLMs, introducing the methods of 'atomic questions' and 'complex factual questions (CFQ)' to further distinguish the defects in visual understanding and knowledge internalization.
SIMS: Simulating Stylized Human-Scene Interactions with Retrieval-Augmented Script Generation
Wenjia Wang (University of Hong Kong), Taku Komura (University of Hong Kong)
GenerationRetrievalTransformerLarge Language ModelReinforcement LearningVideoTextRetrieval-Augmented Generation
🎯 What it does: Proposes the SIMS framework, which combines retrieval-augmented script generation with multi-condition physical control to generate diverse and stylized human-machine-scene interaction animations.
Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation
Fengchen He (Huazhong University of Science and Technology), Shaoqun Zeng (Huazhong University of Science and Technology)
Depth EstimationConvolutional Neural NetworkImage
🎯 What it does: Using ray tracing to simulate dual-pixel (DP) images and train a depth estimation model
Simultaneous Motion And Noise Estimation with Event Cameras
Shintaro Shiba (Keio University), Guillermo Gallego (Technische Universitat Berlin)
Anomaly DetectionOptimizationContrastive LearningOptical FlowVideo
🎯 What it does: A method is proposed to simultaneously estimate motion and noise in event cameras, utilizing event alignment for denoising.
Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras
Petr Hruby (ETH Zurich), Marc Pollefeys (ETH Zurich)
Pose EstimationOptimizationSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes a method for estimating relative pose using the intersection points of linear projections of single-row scan lines in each image from a rolling shutter camera.
SITE: towards Spatial Intelligence Thorough Evaluation
Wenqi Wang (Boston University), Boqing Gong (Boston University)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelImageVideoMultimodalityBenchmark
🎯 What it does: The SITE benchmark has been constructed and released for the systematic evaluation of large visual-language models (VLM) in terms of spatial intelligence (SI) capabilities, covering various contexts including images, videos, static and dynamic scenarios, and perspective changes.
SKALD: Learning-Based Shot Assembly for Coherent Multi-Shot Video Creation
Chen-Yi Lu (Purdue University), Somali Chaterji (Purdue University)
GenerationData SynthesisOptimizationTransformerContrastive LearningVideoText
🎯 What it does: A learning-based multi-shot video stitching method called SKALD is proposed, which automatically selects and arranges candidate shots by calculating the coherence score of video segments to generate coherent short videos.
Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation
Uzay Gökay (University of Bonn), Juergen Gall (University of Bonn)
SegmentationRepresentation LearningConvolutional Neural NetworkAuto EncoderTime SeriesSequential
🎯 What it does: This paper proposes an unsupervised skeletal temporal action segmentation framework called SMQ, which utilizes an encoding-blocking-quantization autoencoder to achieve clustering segmentation of multi-sequence actions.
SketchSplat: 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting
Haiyang Ying (University of Maryland), Matthias Zwicker (University of Maryland)
OptimizationGaussian SplattingMesh
🎯 What it does: Through multi-view differentiable sketch polishing, the optimization of parametric 3D edges is performed directly, avoiding the traditional two-step process of generating point sets and then fitting.
Skip-Vision: Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping
Weili Zeng (Shanghai Jiao Tong University), Yichao Yan (Ant Group)
OptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the Skip-Vision framework, which utilizes Skip-FFN (skipping the FFN computation of visual tokens during the training phase) and Skip-KV-Cache (removing KV cache of skipped tokens during the inference phase) to significantly reduce the training and inference costs of visual language models, while retaining key information through token merge and adaptive summary tokens.
SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing
Yingying Zhang (Ant Group), Jingdong Chen (Ant Group)
ClassificationObject DetectionSegmentationTransformerMixture of ExpertsContrastive LearningImageMultimodality
🎯 What it does: Introducing SkySense V2, a unified Transformer-based multimodal remote sensing foundation model, and introducing the self-supervised pre-training strategy QSACL.
SL2A-INR: Single-Layer Learnable Activation for Implicit Neural Representation
Reza Rezaeian (University of Tehran), Ilker Hacihaliloglu (University of British Columbia)
Image TranslationRestorationSuper ResolutionNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes a SL 2 A-INR that integrates learnable high-order Chebyshev polynomial activation blocks with low-rank ReLU MLP to address the spectral bias problem of traditional INRs.
Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation
Zheyun Qin (Shandong University), Zhumin Chen (Shandong University)
Object DetectionSegmentationTransformerContrastive LearningVideo
🎯 What it does: This paper proposes the SWBridge framework, which achieves cross-domain alignment between the spatial-temporal domain and the text-visual domain through sliced Wasserstein distance, thereby completing the open vocabulary video instance segmentation task.
SliderSpace: Decomposing the Visual Capabilities of Diffusion Models
Rohit Gandikota (Northeastern University), Nick Kolkin (Adobe Research)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: Proposes the SliderSpace framework, which uses unsupervised methods to automatically decompose the visual capabilities of diffusion models into adjustable sliders (control directions).
SMARTIES: Spectrum-Aware Multi-Sensor Auto-Encoder for Remote Sensing Images
Gencer Sumbul (Ecole Polytechnique Federale de Lausanne), Devis Tuia (Ecole Polytechnique Federale de Lausanne)
ClassificationSegmentationDomain AdaptationTransformerAuto EncoderImageMultimodality
🎯 What it does: A unified multi-sensor remote sensing image foundation model, SMARTIES, is proposed, utilizing spectral-aware projection, cross-sensor token mixing, and self-supervised masked autoencoder to achieve shared representation and seamless transfer of data from different sensors.
SMGDiff: Soccer Motion Generation using Diffusion Probabilistic Models
Hongdi Yang (ShanghaiTech University), Lan Xu (ShanghaiTech University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a two-stage real-time user-controllable football motion generation framework called SMGDiff, which combines trajectory generation with autoregressive diffusion models and incorporates contact guidance to generate diverse and high-quality football actions.
SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion
Ahmed Nassar (IBM Research), Peter W. J. Staar (IBM Research)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityTabular
🎯 What it does: We propose SmolDocling, a 256M parameter visual-language model that achieves end-to-end multimodal document conversion, outputting a unified DocTags markup format.
SMoLoRA: Exploring and Defying Dual Catastrophic Forgetting in Continual Visual Instruction Tuning
Ziqi Wang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
ClassificationRecognitionData-Centric LearningTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper addresses the problem of Continuous Visual Instruction Tuning (CVIT) for multimodal large language models and proposes the SMoLoRA framework, which utilizes separable routing (between the visual understanding module and the instruction following module) and adaptive fusion to tackle dual catastrophic forgetting, and constructs a CVIT benchmark covering multiple tasks, multiple instructions, and unseen tasks.
SMP-Attack: Boosting the Transferability of Feature Importance-based Adversarial Attack with Semantics-aware Multi-granularity Patchout
Wen Yang (Chongqing University of Technology), Di Ming (Chongqing University of Technology)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A new transferable adversarial attack method called SMP-Attack is proposed, which integrates semantic-aware multi-granularity patch replacement and multi-stage optimization to enhance the attack success rate of feature importance attacks in black-box scenarios.
SMSTracker: Tri-path Score Mask Sigma Fusion for Multi-Modal Tracking
Sixian Chan, Xiaoqin Zhang
Object TrackingTransformerPrompt EngineeringImageVideoMultimodality
🎯 What it does: Proposes SMSTracker—a three-path multimodal tracking framework that integrates RGB with thermal/depth/event multimodal information, including modules such as Score Mask Fusion, Sigma Interaction, and Drop Key Fine-tuning.
Snakes and Ladders: Two Steps Up for VideoMamba
Hui Lu (Utrecht University), Ronald Poppe (Utrecht University)
RecognitionTransformerVideo
🎯 What it does: This paper proposes VideoMambaPro based on VideoMamba for the video action recognition task. By incorporating masked backward computation and residual connections into the bidirectional state space model, it addresses the issues of historical decay and element contradiction, thereby improving model performance.
Social Debiasing for Fair Multi-modal LLMs
Harry Cheng (Shandong University), Liqiang Nie (Ant Group)
GenerationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageMultimodality
🎯 What it does: This paper constructs a large-scale, multi-concept adversarial dataset called CMSC and proposes the Counter-Stereotype Debiasing (CSD) method to reduce social biases in multimodal large language models (MLLMs).
Soft Local Completeness: Rethinking Completeness in XAI
Ziv Weiss Haddad (Tel Aviv University), Noam Koenigstein (Open University)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: The SLOC method is proposed, which softens the integrity constraints on sub-regions of the attribution map and achieves more trustworthy interpretability maps by minimizing the integrity gap.
Soft Separation and Distillation: Toward Global Uniformity in Federated Unsupervised Learning
Hung-Chieh Fang (National Taiwan University), Yifei Zhang (Chinese University of Hong Kong)
Federated LearningKnowledge DistillationRepresentation LearningContrastive LearningImage
🎯 What it does: Proposes the Soft Separation and Distillation (SSD) framework, which enhances the uniformity and expressiveness of global representations in federated unsupervised learning through dimensional scaling regularization and projector distillation.
SP2T: Sparse Proxy Attention for Dual-stream Point Transformer
Jiaxu Wan (Beijing University of Aeronautics and Astronautics), Yifan Yang (Beijing University of Aeronautics and Astronautics)
Object DetectionSegmentationTransformerPoint Cloud
🎯 What it does: This paper proposes a dual-stream point transformer SP2T based on sparse proxy attention, aimed at achieving a global receptive field in 3D point clouds while maintaining detail extraction.
SPA: Efficient User-Preference Alignment against Uncertainty in Medical Image Segmentation
Jiayuan Zhu (University of Oxford), J. Alison Noble (University of Oxford)
SegmentationTransformerImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes an interactive medical image segmentation framework named SPA, which can efficiently generate multiple representative segmentation results during the inference phase based on user preferences in different clinical contexts and converge quickly.
SPADE: Spatial-Aware Denoising Network for Open-vocabulary Panoptic Scene Graph Generation with Long- and Local-range Context Reasoning
Xin Hu (University of Electronic Science and Technology of China), Tao He (Monash University)
RecognitionSegmentationGenerationGraph Neural NetworkTransformerVision Language ModelDiffusion modelImageGraph
🎯 What it does: Proposes the SPADE framework, which achieves Panoptic Scene Graph Generation under open vocabulary by reverse calibrating UNet through a diffusion model and combining it with a spatially aware relation graph transformer;
Sparfels: Fast Reconstruction from Sparse Unposed Imagery
Shubhendu Jena (Inria), Adnane Boukhayma (Inria)
GenerationOptimizationGaussian SplattingImagePoint Cloud
🎯 What it does: A fast 3D reconstruction method for sparse non-pose images based on the MASt3R pre-trained model and 2D Gaussian Splatting is proposed.
Sparse Fine-Tuning of Transformers for Generative Tasks
Wei Chen (Purdue University), Qiang Qiu (Purdue University)
GenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: This paper proposes viewing the fine-tuning of Transformers as a sparse coding problem, using dictionary atoms and sparse coefficients to achieve interpretable and efficient feature fine-tuning.
Sparse-Dense Side-Tuner for efficient Video Temporal Grounding
David Pujol-Perich (Universitat de Barcelona), Albert Clapés (Universitat de Barcelona)
RecognitionRetrievalTransformerVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: A Sparse-Dense Side Tuning (SDST) framework is proposed for video temporal localization and highlight detection tasks, balancing sparse Moment Retrieval and dense Highlight Detection.
SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling
Xianglong He (Tsinghua University), Yangguang Li (Chinese University of Hong Kong)
GenerationData SynthesisTransformerRectified FlowAuto EncoderPoint CloudMesh
🎯 What it does: We propose SparseFlex, a differentiable isosurface representation with a sparse structure, and build a VAE generation pipeline based on this representation, achieving high-resolution arbitrary topology 3D shape modeling and single-image 3D generation.
SparseLaneSTP: Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection
Maximilian Pittner (Bosch Mobility Solutions), Alexandru Paul Condurache (Bosch Mobility Solutions)
Object DetectionAutonomous DrivingTransformerImage
🎯 What it does: This paper proposes SparseLaneSTP, a 3D lane detection framework that utilizes a sparse Transformer combined with spatial and temporal priors, and constructs a long-distance high-quality lane dataset through automatic labeling techniques.
SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs
Jiahui Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: By analyzing the attention mechanism of multimodal large language models (MLLMs), it reveals that visual-related attention heads are extremely sparse (<5%), and based on this, SparseMM is proposed—a head-level KV-Cache dynamic allocation strategy to achieve efficient acceleration.
SparseRecon: Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies
Liang Han (Tsinghua University), Zhizhong Han (Wayne State University)
RestorationDepth EstimationNeural Radiance FieldPoint CloudBenchmark
🎯 What it does: The SparseRecon method is proposed, which significantly improves the surface reconstruction quality under sparse viewpoints by implementing feature consistency loss through volume rendering and uncertainty-guided depth constraints.
SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference
Samir Khaki (NVIDIA), Zhijian Liu (University of Toronto)
RetrievalComputational EfficiencyTransformerVision Language ModelImageVideo
🎯 What it does: This paper proposes SparseVILA, a framework that decouples visual sparsity during the pre-filling and decoding stages, utilizing query-independent visual token pruning and query-related retrieval to accelerate VLM inference.
Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation
Nairouz Mrabah (ETS Montreal), Eric Granger (ETS Montreal)
Domain AdaptationOptimizationSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a sparse optimization (SO) framework for efficiently fine-tuning visual-language models (VLMs) in scenarios with very few labeled samples.
Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement
Qian Liang (University of Science and Technology of China), Yang Hu (University of Science and Technology of China)
RecognitionTransformerVideoMultimodalityBiomedical Data
🎯 What it does: The SATM adapter is proposed for integrating video and radar data for remote physiological measurement, and a new MMRPM dataset has been collected.
Spatial Preference Rewarding for MLLMs Spatial Understanding
Han Qiu (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
OptimizationTransformerLarge Language ModelReinforcement LearningImageMultimodality
🎯 What it does: By constructing random image regions and utilizing preference optimization to enhance the capability of multimodal large language models (MLLM) for fine-grained spatial understanding.
Spatial-Temporal Aware Visuomotor Diffusion Policy Learning
Zhenyang Liu (Fudan University), Yanwei Fu (Fudan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelGaussian SplattingWorld ModelImagePoint Cloud
🎯 What it does: This paper proposes a 4D diffusion strategy (DP4) that constructs a dynamic Gaussian world model through single-view RGB-D observations, achieving visual imitation learning for 3D spatial and 4D temporal perception.
Spatial-Temporal Forgery Trace based Forgery Image Identification
Yilin Wang (Zhejiang University), Yijun Bei (Zhejiang University)
RecognitionAnomaly DetectionDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A framework for image forgery detection based on spatio-temporal forged traces (STFT) has been developed, utilizing the temporal distribution of latent diffusion models to capture forgery features;
SpatialCrafter: Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations
Songchun Zhang (Zhejiang University), Changqing Zou (Zhejiang University)
GenerationData SynthesisDepth EstimationTransformerDiffusion modelVideo
🎯 What it does: Utilize video diffusion models to generate additional viewpoints, completing 3D scene reconstruction and new viewpoint synthesis based on sparse or single-view input.
Spatially-Varying Autofocus
Yingsi Qin (Carnegie Mellon University), Matthew O'Toole (Carnegie Mellon University)
OptimizationOptical FlowImage
🎯 What it does: An optical system has been developed that can achieve spatially variable focus through phase modulation within the camera, allowing for the direct capture of all-focus images of panoramic scenes without post-processing.
SpatialSplat: Efficient Semantic 3D from Sparse Unposed Images
Yu Sheng (University of Science and Technology of China), Jianmin Ji (University of Science and Technology of China)
SegmentationGenerationTransformerContrastive LearningPoint Cloud
🎯 What it does: A feedforward network is constructed to generate efficient semantic 3D representations from sparse, unposed images, achieving novel view rendering and open vocabulary semantic segmentation.
SpatialTrackerV2: Advancing 3D Point Tracking with Explicit Camera Motion
Yuxi Xiao, Xiaowei Zhou
Object TrackingPose EstimationDepth EstimationTransformerVideo
🎯 What it does: We propose SpatialTrackerV2, an end-to-end, scalable monocular video 3D point tracking method that jointly estimates video depth, camera pose, and pixel-level motion, resulting in consistent 3D scene geometry, camera trajectories, and point trajectories.
Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image
Shin Ishihara (National Institute of Informatics), Imari Sato (National Institute of Informatics)
SegmentationDepth EstimationOptimizationImage
🎯 What it does: A single-spectral image illumination system utilizing diffractive optics is proposed, capable of achieving direct and indirect reflection separation of a scene in a single capture.