arXivSub Start free trial

CVPR 2026 Papers — Page 31

IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers

S$^2$AM3D: Scale-controllable Part Segmentation of 3D Point Clouds

Han Su (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)

SegmentationTransformerContrastive LearningPoint Cloud

🎯 What it does: This paper proposes a framework called S2 AM3D for 3D point cloud part segmentation at any scale.

S2C2Seg: Semantic-Spatial Consistency and Category Optimization for Open-Vocabulary Segmentation

Yuhao Qing (Shanghai University), Xin Xu (National University of Defense Technology)

SegmentationVision Language ModelContrastive LearningImage

🎯 What it does: Proposes an open-vocabulary semantic segmentation framework S2C2Seg without training, leveraging category subset selection (CSS) and consistent semantic guidance (CSG) to enhance segmentation accuracy.

S2D: Selective Spectral Decay for Quantization-Friendly Conditioning of Neural Activations

Arnav Chavan (Amazon), Deepak Gupta (Amazon)

OptimizationComputational EfficiencyTransformerImageMultimodality

🎯 What it does: Investigated the impact of activation outliers on Transformer quantization and proposed the Selective Spectral Decay method for conditioning.

S2D: Sparse to Dense Lifting for 3D Reconstruction with Minimal Inputs

Yuzhou Ji (Shanghai Jiao Tong University), Xin Tan (East China Normal University)

GenerationDiffusion modelGaussian SplattingImagePoint Cloud

🎯 What it does: By combining sparse point clouds with novel view rendering, the sparse input is expanded into a dense point cloud in one step, and a first-order diffusion model is used to repair image distortions. Subsequently, sparse view reconstruction is performed under the 3D Gaussian Splatting (3DGS) framework, achieving high-quality and view-consistent 3D scenes.

S2FT: Parameter-Efficient Fine-Tuning in Sparse Spectrum Domain

Baoquan Zhang (Harbin Institute of Technology Shenzhen), Yao He (Harbin Institute of Technology Shenzhen)

ClassificationGenerationTransformerSupervised Fine-TuningImageTextMultimodality

🎯 What it does: Proposes a framework for parameter-efficient fine-tuning (SFT2) in the sparse spectral domain, utilizing reversible row-column reordering to map roughly estimated weight changes to a sparse spectral latent space, and then adjusting only a few spectral coefficients in this space.

SABER: Spatially Consistent 3D Universal Adversarial Objects for BEV Detectors

Aixuan Li (Northwestern Polytechnical University), Yuchao Dai (Northwestern Polytechnical University)

SegmentationAutonomous DrivingAdversarial AttackTransformerNeural Radiance FieldPoint CloudMesh

🎯 What it does: Proposes a complete framework capable of generating non-intrusive, 3D-consistent adversarial objects to attack BEV-based 3D object detectors.

SafeDrive: Fine-Grained Safety Reasoning for End-to-End Driving in a Sparse World

Jungho Kim (Seoul National University), Jun Won Choi (Seoul National University)

Autonomous DrivingTransformerWorld ModelImagePoint CloudBenchmark

🎯 What it does: SafeDrive proposes an end-to-end driving framework that achieves instance-level safety evaluation by leveraging sparse world models and fine-grained safety reasoning.

SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization

Xuankun Rong (Wuhan University), Mang Ye (Wuhan University)

OptimizationSafty and PrivacyLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelMultimodality

🎯 What it does: Propose a self-reward multimodal safety alignment framework called SafeGRPO based on rule-governed reward, utilizing the GRPO mechanism to optimize the safety of reasoning processes in multimodal large language models.

SafeLogo: Turning Your Logos into Jailbreak Shields via Micro-Regional Adversarial Training

Zhiyi Duan (Inner Mongolia University), Tianxing Man (Inner Mongolia University)

Safty and PrivacyAdversarial AttackLarge Language ModelVision Language ModelMultimodality

🎯 What it does: Proposes SafeLogo, a method that trains logo-sized visual perturbations as a universal defense shield to resist diverse VLM jailbreak attacks while maintaining image quality and model utility.

SafeRoPE: Risk-specific Head-wise Embedding Rotation for Safe Generation in Rectified Flow Transformers

Xiang Yang (Fudan University), Min Yang (Fudan University)

GenerationSafty and PrivacyTransformerFlow-based ModelRectified FlowImageText

🎯 What it does: Proposed a lightweight safe generation framework called SafeRoPE based on RoPE, specifically designed for concept elimination in rectified-flow transformer models such as MMDiT/FLUX;

SAG-GNN: Semantic-Aware Guided GNN for Descriptor-Free 2D-3D Matching

Shihua Zhang (Wuhan University), Jiayi Ma (Wuhan University)

Pose EstimationGraph Neural NetworkVision Language ModelImagePoint Cloud

🎯 What it does: Designed a semantic-aware guided graph neural network, SAG-GNN, for descriptor-free 2D-3D matching, achieving high-precision camera pose estimation.

SAGA: Source Attribution of Generative AI Videos

Rohit Kundu (University of California, Riverside), Amit K. Roy-Chowdhury (YouTube (Google))

ClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningContrastive LearningVideo

🎯 What it does: Proposed the SAGA framework for multi-granularity (authenticity, generation task, Stable Diffusion version, team, specific model) source attribution of AI-generated videos, and provided interpretable Temporal Attention Signatures (T-Sigs);

SAGE: Scalable Agentic 3D Scene Generation for Embodied AI

Hongchi Xia (NVIDIA), Fangyin Wei (NVIDIA)

GenerationData SynthesisRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelDiffusion modelImageTextPoint CloudMesh

🎯 What it does: Built a large-model-based agentic framework called SAGE, capable of automatically generating 3D scenes directly deployable in modern simulators based on open-vocabulary text prompts, supporting multi-level enhancement and action generation for large-scale embodied AI strategy learning.

SAGE: Style-Adaptive Generalization for Privacy-Constrained Semantic Segmentation Across Domains

Qingmei Li, Juepeng Zheng (Sun Yat Sen University)

SegmentationDomain AdaptationSafty and PrivacyTransformerPrompt EngineeringImage

🎯 What it does: A domain generalization framework called SAGE based on input-level visual prompts was designed to improve the performance of semantic segmentation models on unseen target domains when the model is frozen due to privacy or security constraints.

SAGE: Training Smart Any-Horizon Agents for Long Video Reasoning with Reinforcement Learning

Jitesh Jain (SHI Labs Georgia Tech), Humphrey Shi (SHI Labs Georgia Tech)

TransformerReinforcement LearningAgentic AIVision-Language-Action ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation

🎯 What it does: Propose the SAGE system, which utilizes a multi-tool LLM agent to enable video reasoning of arbitrary duration, and trains a flexible inference strategy that can switch between single-step and multi-step reasoning via reinforcement learning.

SAIDO: Generalizable Detection of AI-Generated Images via Scene-Aware and Importance-Guided Dynamic Optimization in Continual Learning

Yongkang Hu (East China Normal University), Zhaoxia Yin (East China Normal University)

Anomaly DetectionTransformerMixture of ExpertsVision Language ModelContrastive LearningImage

🎯 What it does: Propose the SAIDO framework, achieving continuous learning detection for AI-generated images through the Scene-Aware Expert Module (SAEM) and Importance-Guided Dynamic Optimization Mechanism (IDOM).

SAIL: Similarity-Aware Guidance and Inter-Caption Augmentation-based Learning for Weakly-Supervised Dense Video Captioning

Ye-Chan Kim (Hanyang University), Dong-Jin Kim (Hanyang University)

GenerationTransformerLarge Language ModelContrastive LearningVideoText

🎯 What it does: Propose a weakly supervised dense video description method called SAIL, which guides masks through cross-modal similarity and complements sparse annotations using synthetic captions generated by LLMs.

Saliency-Driven Token Merging for Vision Transformers

Weiying Xie (Xidian University), Leyuan Fang (Xidian University)

ClassificationComputational EfficiencyTransformerImage

🎯 What it does: This paper proposes a training-agnostic visual Transformer token merging method called SAD-TM, which determines merging priority by leveraging precomputed visual saliency anomaly scores and class attention during inference.

Saliency-Guided Representation with Consistency Policy Learning for Visual Unsupervised Reinforcement Learning

Jingbo Sun (Chinese Academy of Sciences), Dongbin Zhao (Chinese Academy of Sciences)

Representation LearningReinforcement LearningImageBenchmark

🎯 What it does: Proposed an SRCP framework that leverages significance-guided dynamic representation learning and consistency strategies to achieve zero-shot generalization in visual unsupervised reinforcement learning.

Saliency-R1: Enforcing Interpretable and Faithful Vision-language Reasoning via Saliency-map Alignment Reward

Shizhan Gong (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)

Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality

🎯 What it does: Provide explainable and trustworthy saliency map alignment rewards for the inference process of vision-language models (VLMs), enhancing the interpretability and authenticity of the reasoning.

SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning

Cai Selvas-Sala (Computer Vision Center), Lluis Gomez (Computer Vision Center)

Safty and PrivacyTransformerLarge Language ModelImageTextMultimodalityBenchmark

🎯 What it does: Proposed the SALMUBench benchmark to evaluate the machine forgetting effectiveness of multimodal models such as CLIP when removing sensitive information at the association level.

SAM 3D Body: Robust Full-Body Human Mesh Recovery

Xitong Yang (Meta Superintelligence Labs), Kris Kitani (Meta Superintelligence Labs)

Pose EstimationData-Centric LearningTransformerPrompt EngineeringImageMesh

🎯 What it does: Propose SAM 3D Body (3DB), a promptable single-image full-body 3D human mesh recovery model.

SAM 3D: 3Dfy Anything in Images

Xingyu Chen, Jitendra Malik

Image TranslationGenerationData SynthesisTransformerSupervised Fine-TuningReinforcement LearningFlow-based ModelAuto EncoderGaussian SplattingImageMultimodalityPoint CloudMesh

🎯 What it does: Proposed a model called SAM³D that generates complete 3D object geometry, texture, and layout from a single image, achieving automatic conversion from a single natural image to an interactive 3D scene.

SAM2Text: Towards Prompt-Free and Multi-Resolution Video Scene Text Segmentation

Jing-Yao Zhang (Chinese Academy of Sciences), Fei Yin (Chinese Academy of Sciences)

SegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringImageVideoBenchmark

🎯 What it does: Propose the SAM2Text framework, specializing the Segment Anything Model 2 (SAM2) for video scene text segmentation, achieving automatic segmentation without prompts, high-definition multi-resolution masks, and temporal consistency;

Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination

Zichuan Wang (University of Chinese Academy of Sciences), Jing Dong (Chinese Academy of Science)

Anomaly DetectionExplainability and InterpretabilityVision Language ModelImageTextMultimodality

🎯 What it does: This paper conducts a detailed analysis of the attention distribution and Logit Lens decoding results during the generation process of large vision-language models (LVLMs), revealing the phenomenon of 'same attention, different truths.' It further identifies two types of misreporting mechanisms: visual uncertainty misreporting and context prior misreporting, and proposes a training-agnostic detection-mitigation framework. The framework uses Logit Lens consistency checks to locate misreported objects, followed by targeted mitigation through High Attention Region Masking (HARM) and Visual Evidence Enhanced Decoding (VEED).

Same Content, Different Answers: Cross-Modal Inconsistency in MLLMs

Angela van Sprang (University Of Amsterdam), Yuki M. Asano (University Of Amsterdam)

Explainability and InterpretabilityTransformerMultimodalityBenchmarkChain-of-Thought

🎯 What it does: This paper investigates whether multimodal large language models can maintain consistent reasoning results across different modalities, and proposes two benchmark suites, REST and REST+, for systematic evaluation of cross-modal consistency.

Same or Not? Enhancing Visual Perception in Vision-Language Models

Damiano Marsili, Georgia Gkioxari (California Institute of Technology)

RecognitionTransformerReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodalityBenchmark

🎯 What it does: This paper constructs a large-scale paired visual question answering dataset called TWIN and the corresponding FGVQA fine-grained evaluation suite to train and evaluate visual language models on fine-grained visual perception.

SAME: Sparse and Anchored Model Editing for Heterogeneous Incremental Learning under Limited Data

Zixuan Duan (Nanjing University), Yang Gao (Nanjing University)

Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality

🎯 What it does: Propose a method to achieve continuous learning using sparse and dual-anchored model editing in scenarios with uncertain task boundaries, cross-domain settings, and extremely limited samples.

SAMIX: Reinforcing SAM2 with Semantic Adapter and Reference Selecting Policy for Mix-Supervised Segmentation

Qiang Hu (Huazhong University of Science and Technology), Zhiwei Wang (Huazhong University of Science and Technology)

SegmentationTransformerReinforcement LearningImage

🎯 What it does: Propose the SAMIX framework, which automatically generates high-quality pseudo-labels and trains segmentation models in mix-supervised semantic segmentation by leveraging semantically adapted SAM2 and reinforcement learning-based SPNet.

SAMosaic3D: Modular Scene Assembly for Real-Time 3D Segment Anything

Peng Wang (Renmin University of China), Deying Li (Renmin University of China)

SegmentationComputational EfficiencyConvolutional Neural NetworkTransformerPoint Cloud

🎯 What it does: Developed an online 3D instance segmentation framework called SAMosaic3D, which uses SAM 2D masks as learnable 'puzzle pieces' for block assembly, achieving real-time and coherent 3D instance segmentation.

Sampling-Aware Quantization for Diffusion Models

Qian Zeng (Zhejiang University), Mingli Song (Zhejiang University)

Computational EfficiencyDiffusion modelImage

🎯 What it does: Propose a Sampling-Aware Quantization framework tailored for diffusion models, which can achieve low-bitwidth (e.g., 4-bit, 8-bit) weight/activation quantization while maintaining high generation quality, and also accommodate the sparse steps of high-order samplers;

SAMTok: Representing Any Mask with Two Words

Yikang Zhou (Wuhan University), Xiangtai Li (ByteDance)

SegmentationLarge Language ModelReinforcement LearningImageTextMultimodality

🎯 What it does: Propose SAMTok, which compresses any region mask into two discrete words, enabling multimodal large language models (MLLM) to learn pixel-level capabilities through text prediction

SANER: Switchable Adapter with Non-parametric Enhanced Routing for Person De-Reidentification

Yimin Liu (Hefei University of Technology), Zhun Zhong (Hefei University of Technology)

RecognitionSafty and PrivacyTransformerContrastive LearningImageBenchmark

🎯 What it does: Propose a framework called SANER for human de-identification (De-ReID), which can achieve selective forgetting of forgotten identities while maintaining the accuracy of retained identity recognition;

SaPaVe: Towards Active Perception and Manipulation in Vision-Language Action Models for Robotics

Mengzhen Liu (Peking University), Shanghang Zhang (Peking University)

Robotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelMultimodality

🎯 What it does: Propose SaPaVe, an end-to-end active perception and execution framework based on vision-language-action models, enabling robots to perform active perspective adjustment and grasping/manipulation operations in complex scenarios.

SAQN: Semantic-based Adaptive Query Network for 3D Referring Expression Segmentation

Jiale Huang (University of Science and Technology of China), Shangfei Wang (University of Science and Technology of China)

SegmentationTransformerVision Language ModelContrastive LearningTextMultimodalityPoint Cloud

🎯 What it does: Propose a Semantic-based Adaptive Query Network (SAQN), achieving 3D referential expression segmentation through semantic class queries and an adaptive query fusion module;

SAR2Net: Learning Spatially Anchored Representations for Retrieval-Guided Cross-Stain Alignment

Tianle Shen (Shanghai Jiao Tong University), Xiaofan Zhang (Shanghai Jiao Tong University)

Image HarmonizationRetrievalDomain AdaptationConvolutional Neural NetworkTransformerContrastive LearningBiomedical Data

🎯 What it does: Propose the SAR2Net framework, which models cross-stained whole-image region-level alignment as a retrieval problem by learning spatial anchor-based representations, achieving HE–IHC complementary information alignment without prior registration.

SARL-STG: A Spatially Aware Reinforcement Learning Framework for Refining MLLMs in Spatio-Temporal Video Grounding

Hong Gao (Southeast University), Min-Ling Zhang (Southeast University)

Object DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmark

🎯 What it does: Designed a reinforcement learning-based framework SARL-STG, integrating a pre-trained multimodal large language model (MLLM) for temporal reasoning and an open-source object detector for spatial localization, achieving iterative optimization of spatiotemporal boundaries through multi-stage training.

SARMAE: Masked Autoencoder for SAR Representation Learning

Danxu Liu (Beijing Institute of Technology), Jing Zhang (Wuhan University)

ClassificationObject DetectionSegmentationRepresentation LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningImageMultimodality

🎯 What it does: Propose the SARMAE framework, which leverages self-supervised Masked Autoencoder for representation learning on SAR images and constructs the SAR-1M dataset with optical counterparts at the million-scale.

SASNet: Spatially-Adaptive Sinusoidal Networks for INRs

Haoan Feng (University of Maryland), Leila De Floriani (IMPA)

RestorationSuper ResolutionRepresentation LearningNeural Radiance FieldImageMesh

🎯 What it does: Designed and implemented a spatially adaptive sine network called SASNet to improve frequency control and reconstruction quality in implicit neural representations (INR).

SAT-RRG: LLM-Guided Self-Adaptive Training for Radiology Report Generation with Token-Level Push-Pull Optimization

Yunyi Liu (University of Sydney), Luping Zhou (University of Sydney)

GenerationTransformerLarge Language ModelBiomedical Data

🎯 What it does: Propose the SAT-RRG framework, which utilizes a frozen LLM to mark semantic conflicts in reports during training, and adaptively adjusts token-level gradients through a pull-push loss to enhance report accuracy.

SATTC: Structure-Aware Label-Free Test-Time Calibration for Cross-Subject EEG-to-Image Retrieval

Qunjie Huang (Yunnan University), Weina Zhu (Yunnan University)

RetrievalMixture of ExpertsBiomedical Data

🎯 What it does: This paper proposes SATTC, a head that calibrates the similarity matrix through label-free testing on frozen EEG and image encoders, significantly improving the Top-1/Top-5 results for cross-subject EEG-image retrieval.

SAVA-X: Ego-to-Exo Imitation Error Detection via Scene-Adaptive View Alignment and Bidirectional Cross View Fusion

Xiang Li (University of Electronic Sience and Technology of China), Hongliang Li (University of Electronic Sience and Technology of China)

Anomaly DetectionTransformerVideo

🎯 What it does: Proposed a new Ego → Exo perspective error detection task that can localize erroneous steps in first-person execution within asynchronous videos of different perspectives and lengths.

SAVE: Speech-Aware Video Representation Learning for Video-Text Retrieval

Ruixiang Zhao, Xirong Li (Renmin University Of China)

RetrievalTransformerVision Language ModelContrastive LearningVideoTextMultimodalityAudio

🎯 What it does: Proposed the SAVE method, which employs a three-branch network (visual, audio, speech) with a soft alignment mechanism to enhance video-text retrieval performance.

Say Cheese! Detail-Preserving Portrait Collection Generation via Natural Language Edits

Zelong Sun (Renmin University of China), Zhiwu Lu (Renmin University of China)

GenerationData SynthesisPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality

🎯 What it does: Proposes a novel 'Portrait Collection Generation' (PCG) task, which edits reference portraits through natural language instructions to generate diverse and consistent portrait collections.

Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction

Tao Xie (Zhejiang University), Xiaowei Zhou (Zhejiang University)

OptimizationComputational EfficiencyRepresentation LearningTransformerVideo

🎯 What it does: Proposes the Scal3R framework, which achieves kilometer-scale large-scale 3D reconstruction using only RGB sequences through scalable test-time training and neural global context memory.

Scalable Feature Matching via State Space Modeling and Sparse Correlation

Sin Wai Choo (Northwestern Polytechnical University), Bo Li (Northwestern Polytechnical University)

Pose EstimationDepth EstimationComputational EfficiencyConvolutional Neural NetworkImage

🎯 What it does: Proposed a lightweight and scalable feature matching framework called SLiM, which leverages a Conv-Mamba hybrid backbone, Feature Pyramid Network (FPN), sparse correlation, and recursive coordinate refinement units to achieve low memory and low latency semi-dense matching at high resolution;

Scalable Multi-View Subspace Clustering with Tensorized Anchor Guidance

Miao Jia, Zijian Chen (National University of Defense Technology)

Representation LearningMultimodality

🎯 What it does: This paper proposes an scalable multi-view subspace clustering method called SMVS-TAG, which improves anchor quality through tensorized anchor guidance to achieve large-scale multi-view clustering.

Scalable Object Relation Encoding for Better 3D Spatial Reasoning in Large Language Models

Shengli Zhou (Southern University of Science and Technology), Yang Liu (Peking University)

Representation LearningTransformerLarge Language ModelSupervised Fine-TuningPoint CloudBenchmark

🎯 What it does: Proposes two mechanisms, QuatRoPE and IGRE, to enable more efficient and accurate spatial reasoning for 3D scenes in large language models, and constructs an attribute-free spatial reasoning benchmark (ASR) to evaluate spatial reasoning capabilities independently.

Scalable Trajectory Generation for Whole-Body Mobile Manipulation

Yida Niu (Peking University), Yixin Zhu (Peking University)

Robotic IntelligenceDiffusion modelSequential

🎯 What it does: AutoMoMa, a GPU-accelerated Augmented Kinematic Representation (AKR) planning framework, generates a physically valid trajectory dataset of over 500,000 full-body motion control trajectories.

Scale Space Diffusion

Soumik Mukhopadhyay (University of Maryland), Abhinav Shrivastava (University of Maryland)

GenerationConvolutional Neural NetworkDiffusion modelImage

🎯 What it does: Propose the Scale Space Diffusion (SSD) framework, integrating scale space theory into diffusion models and designing the Flexi-UNet network to achieve multi-resolution reverse diffusion.

Scaling Agentic Reinforcement Learning for Tool-Integrated Reasoning in VLMs

Meng Lu (Virginia Tech), Xuan Wang (Virginia Tech)

Supervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodality

🎯 What it does: Proposed an extensible tool-integrated reinforcement learning training environment called VISTA-Gym, and trained a VLM agent named VISTA-R1 capable of multi-round tool calls and reasoning based on this environment.

Scaling Dense Event-Stream Pretraining from Visual Foundation Models

Zhiwen Chen, Guangming Shi (Xidian University)

SegmentationDepth EstimationKnowledge DistillationRepresentation LearningTransformerOptical FlowMultimodality

🎯 What it does: Proposed a self-supervised cross-modal knowledge distillation framework called ScaleEvent based on Vision Foundation Models (VFM) for large-scale learning of fine-grained event stream representations.

Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset

Qingyan Bai (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)

GenerationData SynthesisVision Language ModelDiffusion modelFlow-based ModelVideoText

🎯 What it does: Proposed the Ditto framework, which constructs a large-scale instruction-based video editing dataset with one million video editing examples through an automated high-quality synthetic data pipeline and a self-supervised instruction generator.

Scaling Multi-Identity Consistency for Image Customization via Multi-to-Multi Matching Paradigm

Yufeng Cheng (ByteDance), Qian He (ByteDance)

RecognitionGenerationReinforcement LearningDiffusion modelContrastive LearningImage

🎯 What it does: Propose a unified multi-identity optimization framework called UMO, which enhances identity consistency in multi-identity image customization through a many-to-many matching paradigm while reducing identity confusion.

Scaling Parallel Sequence Models to Vision Foundation Models

Yitong Jiang (Nvidia), Sifei Liu (Nvidia)

ClassificationObject DetectionSegmentationKnowledge DistillationTransformerVision Language ModelImageText

🎯 What it does: This paper proposes the Compact GSPN (C-GSPN) block, which compresses the 2D propagation of traditional GSPN into a latent space, combines fused CUDA kernels to achieve significant acceleration, and competes in foundational-scale visual pretraining through two-stage cross-operator distillation.

Scaling Self-Supervised and Cross-Modal Pretraining for Volumetric CT Transformers

Cris Claessens (Eindhoven University of Technology), Fons van der Sommen (Eindhoven University of Technology)

ClassificationSegmentationRetrievalTransformerVision Language ModelAuto EncoderContrastive LearningImageTextBiomedical DataComputed Tomography

🎯 What it does: Propose a fully Transformer-based 3D CT base model called SPECTRE, which adopts a two-stage pre-training approach (self-supervised DINO + SigLIP cross-modal alignment) and is trained on public CT scans and radiology reports;

Scaling Spatial Intelligence with Multimodal Foundation Models

Zhongang Cai (SenseTime Research), Lei Yang (SenseTime Research)

Data-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityPoint CloudChain-of-Thought

🎯 What it does: This paper trains multimodal foundation models (InternVL-3, Qwen3-VL, Bagel) on large-scale data to construct the SenseNova-SI-8M dataset, thereby enhancing the model's capabilities in spatial intelligence tasks.

Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework

Kaihua Tang (Tongji University), Jianqiang Huang (Tongji University)

Computational EfficiencyVision Language ModelContrastive LearningMultimodalityBenchmark

🎯 What it does: Propose the Self-Critical Inference (SCI) framework, which enhances the robustness of large vision-language models (LVLMs) through multi-round counterfactual reasoning at both visual and textual levels, and construct a model-specific, dynamic robustness benchmark (DRBench) to evaluate model performance in terms of language bias and language sensitivity.

Scaling the Long Video Understanding of Multimodal Large Language Models via Visual Memory Mechanism

Tao Chen (Xiamen University), Rongrong Ji (Xiamen University)

CompressionComputational EfficiencyLarge Language ModelVision Language ModelVideoMultimodalityRetrieval-Augmented Generation

🎯 What it does: Propose a training-free visual memory mechanism called FlexMem, enabling multimodal large language models (MLLMs) to achieve infinite-length video understanding by iteratively watching videos, writing, and retrieving visual memories.

Scaling Up AI-Generated Image Detection with Generator-Aware Prototypes

Ziheng Qin (Institute of Automation, Chinese Academy of Sciences), Xiaolong Zheng (Institute of Automation, Chinese Academy of Sciences)

Anomaly DetectionTransformerSupervised Fine-TuningVision Language ModelDiffusion modelImageBenchmark

🎯 What it does: Proposes Generator-Aware Prototype Learning (GAPL), which achieves detection of large-scale, diverse AI-generated images by learning prototypes from a few generators and constructing a unified, low-variance feature space through LoRA and attention mapping in two-stage training.

Scaling View Synthesis Transformers

Evan Kim (MIT), Vincent Sitzmann (MIT)

GenerationTransformerImage

🎯 What it does: Propose a scalable view synthesis model SVSM, which achieves computationally optimal view synthesis using an encoder-decoder transformer architecture.

Scaling Zero-Shot Reference-to-Video Generation

Zijian Zhou (Meta AI), Sen He (Meta AI)

GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelAuto EncoderVideoTextMultimodality

🎯 What it does: Propose Saber, a zero-shot reference-video generation framework capable of being trained solely with video-text pairs and achieving high-quality video synthesis from multiple perspectives and multiple references.

Scaling-Aware Data Selection for End-to-End Autonomous Driving Systems

Tolga Dimlioglu (New York University), Jose M. Alvarez (NVIDIA)

Autonomous DrivingOptimizationData-Centric LearningVideo

🎯 What it does: Proposed and implemented a scale-aware iterative sampling framework called MOSAIC, which automatically selects training samples that maximize multi-objective evaluation metrics (e.g., EPDMS) under a limited data budget.

Scaling4D: Pushing the Frontier of Video Novel View Synthesis through Large-Scale Monocular Videos

Hongrui Cai (ByteDance), Qian He (ByteDance)

GenerationData SynthesisTransformerDiffusion modelFlow-based ModelAuto EncoderOptical FlowVideo

🎯 What it does: Propose the Scaling4D framework, which reformulates video novel view synthesis (VNVS) as a pixel correspondence-based generation task. It leverages correspondence information obtained via optical flow for self-supervised training on monocular videos, and enhances diversity through synthetic data.

Scan Clusters, Not Pixels: A Cluster-Centric Paradigm for Efficient Ultra-high-definition Image Restoration

Chen Wu (National University of Defense Technology), Jingyuan Xia (National University of Defense Technology)

RestorationConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderContrastive LearningImage

🎯 What it does: This paper proposes a clustering center scanning-based visual state space model (C2SSM), transforming pixel-level scanning in UHD image restoration into sparse semantic clustering center scanning, significantly reducing computational costs while achieving full-resolution restoration.

SCAPO: Self-Supervised Category-Level Articulated Pose Estimation from a Single 3D Observation

Can Zhang (National University of Singapore), Gim Hee Lee (National University of Singapore)

Pose EstimationAuto EncoderImagePoint Cloud

🎯 What it does: Proposes SCAPO, a self-supervised framework capable of predicting category-level canonical shapes, rigid part segmentation, as well as joint axes, pivots, and motion states from a single RGB-D or point cloud observation, enabling part-level pose and joint parameter estimation with a single observation.

SCE-Depth: A Spherical Compound Eye Framework for Wide FOV Depth Estimation

Yi Zhu (Shanghai Jiao Tong University), Leilei Gu (Shanghai Jiao Tong University)

Depth EstimationConvolutional Neural NetworkTransformerImage

🎯 What it does: Proposes a panoramic depth estimation framework called SCE-Depth based on Spherical Compound Eye (SCE), which directly processes spherical images using a spherical network and fuses gradient features for depth prediction.

SCE-SLAM: Scale-Consistent Monocular SLAM via Scene Coordinate Embeddings

Yuchen Wu (Beihang University), Xiao Bai (Beihang University)

Autonomous DrivingRecurrent Neural NetworkTransformerSimultaneous Localization and MappingOptical FlowImageVideo

🎯 What it does: Proposed an end-to-end monocular visual SLAM system named SCE-SLAM, achieving scale consistency by learning patch-level scene coordinate embeddings;

SceMoS: Scene-Aware 3D Human Motion Synthesis by Planning with Geometry-Grounded Tokens

Anindita Ghosh (DFKI), Rishabh Dabral (Max Planck Institute for Informatics)

GenerationData SynthesisTransformerAuto EncoderImage

🎯 What it does: This paper proposes a text-driven 3D human motion generation framework called SceMoS based on 2D scene information;

ScenDi: 3D-to-2D Scene Diffusion Cascades for Urban Generation

Hanlei Guo (Zhejiang University), Yiyi Liao (Zhejiang University)

GenerationDiffusion modelAuto EncoderGaussian SplattingImageVideoPoint Cloud

🎯 What it does: Propose a ScenDi 3D-to-2D scene diffusion cascade framework: first generate a rough 3D Gaussian scene (including geometry and coarse RGB) using 3D implicit diffusion, then refine details on rendered images with a 2D video diffusion model.

Scene Grounding in the Wild

Tamir Cohen (Tel Aviv University), Hadar Averbuch-Elor (Cornell University)

OptimizationContrastive LearningGaussian SplattingImageMeshBenchmark

🎯 What it does: Align sparse 3D reconstructions derived from internet images with complete pseudo-synthetic reference models to form a globally consistent scene model.

Scene Reconstruction as Mapping Priors for 3D Detection

Yang Fu (Waymo LLC), Yingwei Li (Waymo LLC)

Object DetectionAutonomous DrivingConvolutional Neural NetworkTransformerGaussian SplattingImageMultimodalityPoint Cloud

🎯 What it does: Proposes a 3D object detection framework named MPA3D that integrates scalable scene reconstruction priors (surfels and 3D Gaussians) with LiDAR and camera data fusion.

Scene-Centric Unsupervised Video Panoptic Segmentation

Christoph Reich (TU Munich), Stefan Roth (TU Darmstadt)

SegmentationContrastive LearningOptical FlowVideo

🎯 What it does: Proposed the unsupervised video panoptic segmentation (VPS) task and implemented the VideoCUPS method

Scene-VLM: Multimodal Video Scene Segmentation via Vision-Language Models

Nimrod Berman (Ben Gurion University), Igor Kviatkovsky (Amazon Prime Video)

SegmentationExplainability and InterpretabilityTransformerSupervised Fine-TuningVision Language ModelVideoTextMultimodality

🎯 What it does: Propose SceneVLM, a video scene segmentation framework based on vision-language models (VLM), which progressively predicts consecutive shots by leveraging multimodal information such as visual frames, subtitles, and metadata;

SceneMaker: Open-set 3D Scene Generation with Decoupled De-occlusion and Pose Estimation Model

Yukai Shi (Tsinghua University), Lei Zhang (IDEA Research)

GenerationPose EstimationSupervised Fine-TuningDiffusion modelFlow-based ModelImagePoint CloudMesh

🎯 What it does: Proposed SceneMaker, a decoupled 3D scene generation framework capable of generating high-quality open 3D scenes from a single image.

Scenes as Tokens: Multi-Scale Normal Distributions Transform Tokenizer for General 3D Vision-Language Understanding

Yutao Tang (Johns Hopkins University), Mei Chen (Johns Hopkins University)

RecognitionSegmentationCompressionRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningTextMultimodalityPoint Cloud

🎯 What it does: Designed and implemented NDTokenizer3D, a multi-scale NDT (Normal Distributions Transform) converter that compresses high-resolution point clouds into information-rich scene tokens, supporting multi-task applications such as 3D visual question answering, dense description, referring expression segmentation, and being compatible with human-computer interaction prompts;

SceneScribe-1M: A Large-Scale Video Dataset with Comprehensive Geometric and Semantic Annotations

Yunnan Wang (Shanghai Jiao Tong University), Yujun Shen (Ant Group)

Object TrackingGenerationDepth EstimationLarge Language ModelVision Language ModelVideoMultimodalityBenchmark

🎯 What it does: Constructed and publicly released SceneScribe-1M, a large-scale multimodal video dataset containing approximately one million real-world videos, 4,000+ hours of duration, comprehensive geometric annotations (camera poses, continuous depth, 3D point trajectories), and semantic annotations (structured descriptions, dynamic masks), and evaluated on multiple benchmark tasks including depth estimation, scene reconstruction, dynamic point tracking, and text/pose-driven video generation.

SceneTok: A Compressed, Diffusable Token Space for 3D Scenes

Mohammad Asim (Max Planck Institute for Informatics, Saarland Informatics Campus), Jan Eric Lenssen (Max Planck Institute for Informatics, Saarland Informatics Campus)

GenerationCompressionTransformerDiffusion modelRectified FlowNeural Radiance FieldAuto EncoderImage

🎯 What it does: Propose SceneTok, an autoencoder that compresses multi-view images into a small set of unstructured, diffusable tokens, and utilizes a lightweight generative decoder to render 3D scenes from new trajectories.

SciEducator: Scientific Video Understanding and Educating via Deming-Cycle Multi-Agent System

Zhiyu Xu, Hehe Fan (Zhejiang University)

TransformerLarge Language ModelAgentic AIVision-Language-Action ModelVideoMultimodalityBenchmarkPhysics RelatedRetrieval-Augmented Generation

🎯 What it does: Designed and implemented SciEducator, an iterative self-evolving multi-agent system based on the Deming cycle for scientific video understanding and educational content generation.

SCIEval: Evaluating and Benchmarking the Faithfulness of Scientific Image Generation and Interpretation with Large Multimodal Models

Guanghui Ye (Hunan University), Zhihua Jiang (Jinan University)

GenerationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningMultimodalityBenchmark

🎯 What it does: Proposed a SCIEval framework for interpretable credibility evaluation of scientific image generation and explanation, covering three dimensions: relevance, accuracy, and interpretability.

Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling

Yuran Wang (Peking University), Wentao Zhang (Peking University)

GenerationTransformerMixture of ExpertsImageBenchmark

🎯 What it does: Proposed the Scone model, combining a unified understanding-generation architecture to achieve multi-agent composition and agent distinction, and designed a two-phase training process with an understanding bridge strategy; simultaneously released the SconeEval benchmark for evaluating composition and distinction capabilities in multi-agent image generation.

SCoRe: Salience-Coverage Reduction for Vision Token Pruning in Vision-Language Models

Tong Xu (Institute of Microelectronics, Chinese Academy of Sciences), Xingyu Gao (Institute of Microelectronics, Chinese Academy of Sciences)

OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoMultimodality

🎯 What it does: Propose a unified representative optimization framework called SCoRe based on the weighted k-center problem for pruning visual tokens in vision-language models.

Score2Instruct: Scaling Up Video Quality-Centric Instructions via Automated Dimension Scoring

Qizhi Xie (Tsinghua University), Jihong Zhu (Kuaishou Technology)

Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought

🎯 What it does: Constructed a large-scale automatically generated video quality assessment instruction dataset, and used this dataset to progressively fine-tune large multimodal models, enabling them to simultaneously perform video quality scoring and textual explanation.

Sculpt4D: Generating 4D Shapes via Sparse-Attention Diffusion Transformers

Minghao Yin (University of Hong Kong), Kai Han (University of Hong Kong)

GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelVideoMesh

🎯 What it does: Designed a model called Sculpt4D that can generate high-fidelity 4D mesh sequences from videos.

SD-FSMIS: Adapting Stable Diffusion for Few-Shot Medical Image Segmentation

Meihua Li (Shenzhen University), Yisong Li (Shenzhen University)

SegmentationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography

🎯 What it does: This paper proposes a few-shot medical image segmentation framework called SD-FSMIS based on Stable Diffusion, leveraging the general visual prior of pre-trained diffusion models to accomplish few-shot segmentation tasks.

SDDF: Specificity-Driven Dynamic Focusing for Open-Vocabulary Camouflaged Object Detection

Jiaming Liang (Shenzhen University), Qiang Nie (Hong Kong University of Science and Technology (Guangzhou))

Object DetectionConvolutional Neural NetworkVision Language ModelContrastive LearningImageTextMultimodalityBenchmark

🎯 What it does: This paper proposes the open-vocabulary camouflaged object detection (OVCOD) task, constructs the OVCOD-D dataset, and designs the SDDF method based on visual-language pre-trained models to specifically address the challenge of detecting camouflaged objects when they are highly visually similar to the background.

SDGS: Spatial Difference Guided Gaussian Splatting for Simultaneous Localization and 3D Reconstruction

Yijian Tian (Tsinghua University), Xinglong Ji (Tsinghua University)

Pose EstimationAutonomous DrivingGaussian SplattingSimultaneous Localization and MappingOptical FlowImage

🎯 What it does: Propose a 3D Gaussian splitting online SLAM framework based on sparse spatial difference (SD) edges, which can simultaneously achieve camera 6-DoF localization and dense 3D reconstruction without prior poses;

SDTrack: A Baseline for Event-based Tracking via Spiking Neural Networks

Yimeng Shan (University of Electronic Science and Technology of China), Malu Zhang (University of California, Santa Cruz)

Object TrackingSpiking Neural NetworkTransformerSupervised Fine-TuningTime Series

🎯 What it does: Proposes an end-to-end event camera single-target tracking framework SDTrack, utilizing GTP event aggregation and a full Spiking Neural Network (SNN) tracker based on Transformer.

SDUIE: Semi-Supervised Diffusion for Underwater Image Enhancement with Quant-Text Dual Control

Xiaofeng Cong (Southeast University), Jie Gui (Southeast University)

RestorationConvolutional Neural NetworkPrompt EngineeringVision Language ModelDiffusion modelImageMultimodality

🎯 What it does: Propose a semi-supervised diffusion framework SDUIE that achieves underwater image enhancement through dual methods of numerical control and text prompts, supporting seamless migration from synthetic to real scenarios.

SE(3)-Equivariance with Geometric and Topological Guidance for Category-Level Object Pose Estimation

Sheng Yu (Beijing Institute of Technology), Yuanqing Xia (Beijing Institute of Technology)

Pose EstimationGraph Neural NetworkTransformerPoint Cloud

🎯 What it does: Propose SEGPose, a method for category-level 6D object pose estimation using only depth point clouds.

SEA-Flow3D: Simplified, Efficient, and Accurate Scene Flow via Spatial Vector Sampling and Multi-scale Refinement

Han Ling (Nanjing University of Science and Technology), Yinghui Sun (Southeast University)

Autonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowImageBenchmark

🎯 What it does: Propose a dense scene flow estimation framework SEA-Flow3D based on RGB-Depth, achieving precise recovery of pixel-level 3D motion by simultaneously leveraging 2D feature correlations and 3D spatial orientation information during iterative optimization.

SEA-Vision: A Multilingual Benchmark for Comprehensive Document and Scene Text Understanding in Southeast Asia

Pengfei Yue (Xiamen University), Liujuan Cao (Xiamen University)

MultimodalityBenchmark

🎯 What it does: Propose SEA-Vision, a unified multilingual benchmark for evaluating document parsing and text-centric visual question answering, covering 11 Southeast Asian languages.

SEA: Evaluating Sketch Abstraction Efficiency via Element-level Commonsense Visual Question Answering

Jiho Park (Dongguk University), Jihie Kim (Dongguk University)

Large Language ModelVision Language ModelImageTextBenchmark

🎯 What it does: Propose a no-reference sketch abstraction efficiency evaluation metric SEA, which utilizes common-sense visual elements and a visual question-answering model to quantify the balance between maintaining recognizability and visual simplicity in sketches.

SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models

Jiwoo Chung (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)

GenerationComputational EfficiencyDiffusion modelImageVideo

🎯 What it does: SeaCache introduces a cache strategy based on spectral evolution during the sampling process of diffusion models, dynamically deciding when to reuse intermediate features to accelerate inference.

SearchAD: Large-Scale Rare Image Retrieval Dataset for Autonomous Driving

Felix Embacher (Mercedes-Benz AG), Markus Enzweiler (Esslingen University of Applied Sciences)

RetrievalAutonomous DrivingVision Language ModelImageMultimodalityBenchmark

🎯 What it does: Constructed a large-scale rare image retrieval dataset named SearchAD, focusing on extremely rare objects and scenes in autonomous driving;

SEASON: Mitigating Temporal Hallucination in Video Large Language Models via Self-Diagnostic Contrastive Decoding

Chang-Hsun Wu (National Taiwan University), Yu-Chiang Frank Wang (NVIDIA)

TransformerLarge Language ModelVision Language ModelContrastive LearningVideoMultimodalityBenchmark

🎯 What it does: Propose a training-free self-diagnostic contrastive decoding framework named SEASON, aiming to eliminate temporal and spatial hallucinations in video large language models (VideoLLM), thereby enhancing the authenticity and reliability of video understanding.

SEATrack: Simple, Efficient, and Adaptive Multimodal Tracker

Junbin Su (Yanshan University), Zhipeng Zhang (Shanghai Jiao Tong University)

Object TrackingComputational EfficiencyTransformerMixture of ExpertsMultimodality

🎯 What it does: Propose SEATrack, a simple, efficient, and adaptive two-stream multi-modal tracker that addresses the contradiction between high parameter count and low performance.

SEBA: Sample-Efficient Black-Box Attacks on Visual Reinforcement Learning

Tairan Huang (Hong Kong Polytechnic University), Haibo Hu (Hong Kong Polytechnic University)

Adversarial AttackTransformerReinforcement LearningGenerative Adversarial NetworkWorld ModelImage

🎯 What it does: Propose the SEBA framework to achieve black-box adversarial attacks in visual reinforcement learning

SECOS: Semantic Capture for Rigorous Classification in Open-World Semi-Supervised Learning

Hezhao Liu (Xiamen University), Yang Lu (Xiamen University)

ClassificationTransformerPrompt EngineeringMultimodality

🎯 What it does: Propose the SECOS framework to address strict classification (RC-OWSSL) in open-world semi-supervised learning, enabling direct prediction of candidate text labels without post-processing.

SeD-UD: An Influence-Driven and Hierarchically-Decoupled Information Bottleneck for Multimodal Intent Recognition

Qin Li (Hunan University of Technology and Business), Guanying Xu (Hunan First Normal University)

RecognitionRepresentation LearningMultimodalityBenchmark

🎯 What it does: This paper proposes a multi-modal intent recognition framework named SeD-UD based on the information bottleneck theory, which integrates an IDAB module with adaptive input scheduling to achieve hierarchical purification of redundant information and noise.