CVPR 2026 Papers with Code — Page 8

RaPA: Enhancing Transferable Targeted Attacks via Random Parameter Pruning

Tongrui Su (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)

🎯 What it does: Studied a Random Parameter Pruning Attack (RaPA), which enhances the success rate of transferable targeted attacks by randomly pruning parameters of the substitute model during the attack process.

RARE: Learn to RAnk and REtrieve for Monocular 3D Object Detection

Hyeonjeong Park (University of Illinois Chicago), Wei Tang (University of Illinois Chicago)

🎯 What it does: Propose a unified Rank-and-Retrieve framework for monocular 3D object detection, improving confidence estimation and achieving multi-modal 3D localization.

RAVEN: Erasing Invisible Watermarks via Novel View Synthesis

Fahad Shamshad (MBZUAI), Karthik Nandakumar (MBZUAI)

🎯 What it does: The study proposes a view synthesis attack based on a zero-shot diffusion model, capable of erasing invisible watermarks without accessing watermark information or models.

RDFace: A Benchmark Dataset for Rare Disease Facial Image Analysis under Extreme Data Scarcity and Phenotype-Aware Synthetic Generation

Ganlin Feng (Western University), Pingzhao Hu (Western University)

🎯 What it does: Constructed the RDFace benchmark dataset of facial images for rare diseases, and conducted experiments based on supervised learning, few-shot learning, and synthetic data augmentation using generative models, evaluating the diagnostic performance of different models under extremely low sample conditions.

Re-evaluating Continual VQA: Toward Fair and Robust Evaluation for Multimodal Continual Learning

Zijian Gao (National University of Defense Technology), Huaimin Wang (National University of Defense Technology)

🎯 What it does: Proposes a fairer and more robust multi-modal continual visual question answering (Continual VQA) evaluation framework, UCo-VQA, and designs a lightweight continual learning method, MaDQ, which enhances knowledge retention and visual-semantic alignment through question replay and dual-layer distillation;

RE-VLM: Event-Augmented Vision-Language Model for Scene Understanding

Hanqing Liu (Beijing University of Posts and Telecommunications), Chuang Zhu (Beijing University of Posts and Telecommunications)

🎯 What it does: Proposed RE-VLM, a dual-stream vision-language model that integrates RGB images and event streams, enabling robust scene understanding under adverse conditions such as poor illumination or high-speed motion.

Reading or Reasoning? Format Decoupled Reinforcement Learning for Document OCR

Yufeng Zhong (Meituan), Lin Ma (Meituan)

🎯 What it does: Proposed and implemented a document OCR framework based on format-decoupled reinforcement learning (FD-RL), achieving high-quality end-to-end recognition of text, formulas, and tables through two-stage SFT+RL training.

ReaGEN: Adaptive Generation of Structured Chains-of-Thought for Efficient Multimodal Reasoning

Ruiqing Tian (Huawei Technologies), Bahador Rashidi (Huawei Technologies)

🎯 What it does: Proposes the ReaGEN framework, which acquires sample-specific structured chain-of-thought (CoT) through teacher-guided evolutionary search, and trains a lightweight generator (GEN) to generate adaptive reasoning chains without altering the underlying vision-language model, thereby enhancing multimodal reasoning performance.

Reallocating Attention Across Layers to Reduce Multimodal Hallucination

Haolang Lu (Beijing University of Posts and Telecommunications), Kun Wang (Nanyang Technological University)

🎯 What it does: Propose a lightweight, training-agnostic plugin that first identifies shallow perceptual heads and deep reasoning heads via a functional head, then applies category-conditioned rescaling to them, thereby mitigating hallucination issues in multi-modal large reasoning models without altering the model architecture.

RealUnify: Do Unified Models Truly Benefit from Unification? A Comprehensive Benchmark

Yang Shi (Peking University), Ziwei Liu (Nanyang Technological University)

🎯 What it does: This paper proposes the RealUnify benchmark for systematically evaluating the bidirectional synergy between two core capabilities of unified models: visual understanding and generation;

ReasonEdit: Towards Reasoning-Enhanced Image Editing Models

Fukun Yin (StepFun), Daxin Jiang (StepFun)

🎯 What it does: Proposed a novel image editing framework called ReasonEdit, which incorporates dual reasoning mechanisms of Thinking and Reflection, enabling dynamic parsing of abstract instructions and self-correction during the editing process.

Reasoning-Driven Anomaly Detection and Localization with Image-Level Supervision

Yizhou Jin (Beihang University), Yunhong Wang (Beihang University)

🎯 What it does: This paper proposes an end-to-end industrial anomaly detection framework based on a multimodal large language model (MLLM), which realizes anomaly detection, pixel-level localization, and interpretable reasoning through the model's own reasoning process, without relying on external visual modules or pixel-level annotations.

ReasonX: MLLM-Guided Intrinsic Image Decomposition

Alara Dirik (Imperial College London), Anna Frühstück (Adobe Research)

🎯 What it does: Propose the ReasonX framework, which evaluates relative intrinsic image decomposition through multimodal large language models (MLLM) and performs fine-tuning on unlabeled real images using Group Relative Policy Optimization (GRPO).

ReCALL: Recalibrating Capability Degradation for MLLM-based Composed Image Retrieval

Tianyu Yang (Chinese Academy of Sciences), Tat-Seng Chua (National University of Singapore)

🎯 What it does: Proposes a diagnosis-generation-correction framework named ReCALL to address the capability degradation issue when migrating generative multi-modal large language models (MLLM) to retrieval models;

RecEdit-Drive: 3D Reconstruction-Guided Spatiotemporal Video Editing for Autonomous Driving Scenes

Yipeng Wu (Tianjin University), Di Lin (Tianjin University)

🎯 What it does: Propose the RecEdit-Drive framework for achieving high-quality, spatiotemporally consistent foreground object editing (deletion, replacement, insertion, relocation) in autonomous driving scenarios

Reclaiming Lost Text Layers for Source-Free Cross-Domain Few-Shot Learning

Zhenyu Zhang (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)

🎯 What it does: This paper investigates the phenomenon of 'lost layers' in the CLIP text encoder within the source-free cross-domain few-shot learning (SFCDFSL) task, where certain intermediate layers are considered redundant but actually contribute to performance. It proposes a model named VtT, which reutilizes these lost layers through hierarchical fusion, text information absorption, and dynamic gradient supervision to enhance model performance in cross-domain few-shot tasks.

Recover to Predict: Progressive Retrospective Learning for Variable-Length Trajectory Prediction

Hao Zhou (Great Bay University), Fei Luo (Great Bay University)

🎯 What it does: Propose a Progressive Retrospective Framework (PRF) that gradually maps variable-length, missing trajectories to standard complete trajectories through multi-level recursive units, achieving high-precision prediction for incomplete observations.

RECS4R: Bridging Semantics and Geometry for Referring Remote Sensing Interpretation

Jinming Chai (Xidian University), Weibin Li (Xidian University)

🎯 What it does: This paper proposes RECS4R, a unified multi-task framework for visual grounding (VG) and referential semantic segmentation (RIS) in remote sensing images, achieving precise localization and segmentation through language-guided unified contour decoding, residual coarse-to-fine fusion, channel-isolated multi-scale fusion, and gradient consistency loss.

Refacade: Editing Object with Given Reference Texture

Youze Huang (University of Electronic Science and Technology of China), Rong Xiao (IntelliFusion Inc.)

🎯 What it does: Propose Refacade, addressing the object retexturing task in videos, which can precisely transfer textures from reference images to target objects while preserving the target's geometric structure.

Refining Few-Step Text-to-Multiview Diffusion via Reinforcement Learning

Ziyi Zhang (Wuhan University), Lefei Zhang (Wuhan University)

🎯 What it does: Propose MVC-ZigAL, a reinforcement learning fine-tuning framework for few-step text-to-multi-view diffusion models;

ReflexSplit: Single Image Reflection Separation via Layer Fusion-Separation

Chia-Ming Lee (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)

🎯 What it does: This paper proposes a dual-stream framework called ReflexSplit for single image reflection separation, decomposing the mixed image into transmission and reflection layers.

ReFTA: Breaking the Weight Reconstruction Bottleneck in Tensorized Parameter-Efficient Fine-Tuning

Jingjing Zheng (Peking University), Yankai Cao (University of British Columbia)

🎯 What it does: This paper proposes an unweighted tensor parameter-efficient fine-tuning method called ReFTA;

RefTon: Reference person shot assist virtual Try-on

Liuzhuozheng Li (University of Tokyo), Yuhui Yin (360 AI Research)

🎯 What it does: Propose a RefTon framework based on Flux, achieving virtual try-on using only the source person image and target clothing image, and enhancing texture details and realism through additional reference images (photos of the same clothing worn by others).

ReGenHOI: Unifying Reconstruction and Generation for 3D Human-Object Interaction Understanding

Miao Xu (Chinese Academy of Sciences), Zhen Lei (Chinese Academy of Sciences)

🎯 What it does: This paper proposes a unified framework that can reconstruct 3D human-object interaction states from images and generate reasonable interaction actions based on natural language.

RegFormer: Transferable Relational Grounding for Efficient Weakly-Supervised Human-Object Interaction Detection

Jihwan Park (KAIST), Hyunwoo J. Kim (KAIST)

🎯 What it does: Achieved weakly supervised human-object interaction detection (HOI) using only image-level labels for training, enabling direct inference of instance-level interaction information.

RegionFuse: Region-Adaptive Pixel Distribution Learning for Infrared and Visible Image Fusion

Jianghan Xia (Beijing Institute of Technology), Jian Yang (Beijing Institute of Technology)

🎯 What it does: This paper proposes RegionFuse, a region-adaptive fusion network based on local pixel distribution, for infrared and visible light image fusion.

Rejection Mixing: Fast Semantic Propagation of Mask Tokens for Efficient DLLM Inference

Yushi Ye (Shanghai Jiao Tong University), Jiangchao Yao (Shanghai Jiao Tong University)

🎯 What it does: This paper proposes a training-agnostic decoding framework called ReMix, which introduces a Continuous Mixing State between Mask and Token states and applies a Rejection Rule to iteratively refine combinatorial contradictions in parallel decoding of DLLMs, significantly improving inference speed while maintaining or even enhancing generation quality.

ReLaX: Reasoning with Latent Exploration for Large Reasoning Models

Shimin Zhang (Hong Kong Polytechnic University), Jibin Wu (Hong Kong Polytechnic University)

🎯 What it does: Propose ReLaX, an RLVR training framework that utilizes Koopman transform to analyze the hidden state dynamics of large reasoning models (LRMs), employing Dynamic Spectral Dispersion (DSD) to measure the diversity of hidden state dynamics and embedding it into the GRPO loss to regulate the exploration-exploitation balance.

Reliable Policy Transfer for Safety-Aware End-to-End Driving with Deep Reinforcement Learning

Uddin Md. Borhan (Shenzhen University), Jie Chen (Shenzhen University)

🎯 What it does: Propose a unified deep reinforcement learning framework that realizes safety-aware end-to-end autonomous driving through a control layer reliability interface.

ReMatch: Boosting Representation through Matching for Multimodal Retrieval

Qianying Liu (University of Glasgow), Paul Henderson (Xiaohongshu Inc.)

🎯 What it does: ReMatch achieves end-to-end representation learning for image-text retrieval tasks by introducing a chat-style generative matching phase and multi-learnable tokens for multi-vector fusion on multimodal large language models (MLLM).

RemedyGS: Defend 3D Gaussian Splatting Against Computation Cost Attacks

Remote Sensing Image Super-Resolution for Imbalanced Textures: A Texture-Aware Diffusion Framework

Enzhuo Zhang (Nanjing University), Pengfeng Xiao (Nanjing University)

🎯 What it does: This paper proposes the TexADiff framework, addressing the texture imbalance problem in remote sensing image super-resolution by using the texture relative density map (RTDM) for spatial conditioning, loss weighting, and sampling scheduling.

Representation-Steered Incremental Adapter-Tuning for Class-Incremental Learning with Pre-Trained Models

Jiarui Zhao (Chinese Academy of Sciences), Yongjun Xu (Chinese Academy of Sciences)

🎯 What it does: Proposes Representation-Steered Incremental Adapter Tuning (RSIAT), a class-incremental learning framework based on pre-trained models, which employs shared adapters and achieves stable and efficient incremental learning through representation-guided loss, residual autoencoder projection, and orthogonal loss.

Repurposing 3D Generative Model for Autoregressive Layout Generation

Haoran Feng (Beihang University), Lu Sheng (Beihang University)

🎯 What it does: Propose the LaviGen framework, utilizing 3D generative models to achieve autoregressive native 3D layout generation;

ReSAM: Refine, Requery, and Reinforce: Self-Prompting Point-Supervised Segmentation for Remote Sensing Images

Muhammad Naseer Subhani (Independent Researcher)

🎯 What it does: Proposed a point-supervised self-prompting framework called ReSAM, which gradually converts sparse point annotations into high-quality pseudo masks through a Refine-Requery-Reinforce cycle, achieving unsupervised domain adaptation for remote sensing images.

Residual Connections Harm Generative Representation Learning

Xiao Zhang (University of Chicago), Michael Maire (University of Chicago)

🎯 What it does: Introduce layer-wise decayed identity shortcuts into self-supervised generative models (Masked Autoencoders and Diffusion Models) to reduce the direct transmission of low-level features in residual networks, thereby enhancing semantic feature learning and generation quality.

Residual Diffusion Bridge Model for Image Restoration

Hebaixu Wang (Wuhan University), Bo Du (Wuhan University)

🎯 What it does: Propose the Residual Diffusion Bridge Model (RDBM) for unified handling of various image restoration tasks.

Resolving Endpoint Underfitting in Diffusion Bridges via Noise Alignment

Yurong Gao (University of Chinese Academy of Sciences), Xinmin Qiu (University of Chinese Academy of Sciences)

🎯 What it does: This paper proposes the Noise-Aligned Diffusion Bridge (NADB), addressing the underfitting issue at the target endpoint of existing diffusion bridges by redesigning the noise-aligned bridge model.

Resolving Evidence Sparsity: Agentic Context Engineering for Long-Document Understanding

Keliang Liu (Fudan University), Lihua Zhang (Fudan University)

🎯 What it does: Proposes SLEUTH, a no-training, plug-and-play multi-agent framework that constructs highly dense, cross-modal evidence context through a coarse-to-fine process, utilizing modules such as retrieval, clue discovery, page screening, difficulty assessment, and core decision-making, to achieve long document question answering.

Resolving the Stability-Plasticity Dilemma in Reinforcement Learning via Complementary Continual Critics

Bo Sun (Sun Yat-sen University), Luntong Li (Peng Cheng Laboratory)

🎯 What it does: Proposed a Continual Dual-Critic with Cross-Attention (CD-CCA) framework that leverages dual Critics combined with Continual Backpropagation (CBP) and Elastic Weight Consolidation (EWC), and dynamically fuses value estimates through Cross-Attention to address the plasticity-stability dilemma in visual reinforcement learning.

Rethinking Concept Bottleneck Models: From Pitfalls to Solutions

Merve Tapli (Middle East Technical University), Emre Akbas (Middle East Technical University)

🎯 What it does: Propose the CBM-Suite method, systematically addressing four major defects of concept bottleneck models: concept relevance assessment, linear problems, accuracy gap, and selection of backbones and VLMs.

Rethinking Diffusion Model-Based Video Super-Resolution: Leveraging Dense Guidance from Aligned Features

Jingyi Xu (Beihang University), Mai Xu (Beihang University)

🎯 What it does: Propose DGAF-VSR, which combines diffusion models with a densely feature-guided video super-resolution network;

Rethinking Glyph Spatial Information in Font Generation

Peng Su (Jilin University), Xi Yang (Jilin University)

🎯 What it does: Proposed the Spatial Preservation Rendering (SPR) scheme, constructed a large-scale OFL-licensed Chinese font dataset, and designed a two-stage GlyphSpatialNet model to achieve end-to-end generation of high-quality vector fonts from reference fonts.

Rethinking Knowledge Transfer in Image Quality Assessment: A Perceptual Preference Structure Alignment Perspective

Aobo Li (Xidian University), Weisheng Dong (Xidian University)

🎯 What it does: This paper proposes a knowledge transfer framework called PreSTA based on perceptual preference structure alignment, aiming to enhance the generalization performance of image quality assessment models across different scenarios.

Rethinking SNN Online Training and Deployment: Gradient-Coherent Learning via Hybrid-Driven LIF Model

Zecheng Hao (Peking University), Tiejun Huang (Peking University)

🎯 What it does: Proposed the Hybrid-Driven Leaky-Integrate-and-Fire (HD-LIF) model family, achieving gradient consistency in online learning and improving training and inference efficiency through various variants.

Rethinking UMM Visual Generation: Masked Modeling for Efficient Image-Only Pre-training

Peng Sun (Westlake University), Tao Lin (Westlake University)

🎯 What it does: Train the visual generation module of a unified multimodal model, proposing a two-stage image-only pre-training and hybrid fine-tuning framework called IOMM.

Retrieve-to-Restore: Efficient All-in-One Image Restoration with a Retrieval-Based Degradation Bank

Chenxu Wang (Nanjing University), Jian Yang (Nanjing University)

🎯 What it does: Propose an All-in-One image restoration framework R2R, which externalizes degradation information by leveraging a retrieval-based degradation knowledge base, enabling a single model to simultaneously handle multiple degradations.

RevINN: An End-to-End Invertible Neural Network for Reversible Adversarial Examples Generation

Jielun Huang (University of Macau), Guoheng Huang (Guangdong University of Technology)

🎯 What it does: Proposed an end-to-end reversible neural network called RevINN, which can directly generate reversible adversarial examples (RAE) from the frequency information of images and restore the original image during the reverse process.

Revisiting Learning with Noisy Labels: Active Forgetting and Noise Suppression

Mengmeng Sheng (Nanjing University of Science and Technology), Fumin Shen (University of Electronic Science and Technology of China)

🎯 What it does: This paper proposes a framework called FINE for learning tasks with noisy labels.

Revisiting Token Compression for Accelerating ViT-based Sparse Multi-View 3D Object Detectors

Mingqian Ji (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)

🎯 What it does: This paper proposes the SEPatch3D framework, which accelerates sparse multi-view 3D object detection models based on Vision Transformer through dynamic patch size selection, information patch screening, and cross-granularity feature enhancement.

RGB-Event based Pedestrian Attribute Recognition: A Benchmark Dataset and An Asymmetric RWKV Fusion Framework

Xiao Wang (Anhui University), Chenglong Li (Anhui University)

🎯 What it does: Constructed the first multi-modal RGB-Event pedestrian attribute recognition dataset, EventPAR, and proposed an RWKV-based heterogeneous RGB-Event fusion framework for joint recognition of pedestrian appearance attributes and emotional attributes.

RHCNet: Residual-Guided Hierarchical Calibration Network for Robust Underwater Object Detection

Yueying Wang (Shanghai University), Xin Xu (East China University of Science and Technology)

🎯 What it does: Proposed a residual-guided hierarchical calibration network (RHCNet), achieving structural restoration, semantic focusing, and cross-scale alignment for blurred, low-contrast targets in seawater environments through the residual-guided feature enhancement module (RGFE) and hierarchical feature calibration pyramid (HFCP).

RHO: Robust Holistic OSM-Based Metric Cross-View Geo-Localization

Junwei Zheng (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)

🎯 What it does: Proposed a robust metric cross-view geolocation framework called RHO for OSM maps, and created a large-scale robust dataset named CV-RHO.

RiskProp: Collision-Anchored Self-Supervised Risk Propagation For Early Accident Anticipation

Yiyang Zou (Wuhan University), Yu Wu (Wuhan University)

🎯 What it does: Propose a self-supervised risk propagation framework called RiskProp that uses only collision frames as supervision for early prediction of collisions.

RMIR: A Benchmark Dataset for Reasoning-Intensive Multimodal Image Retrieval

Yijiang Li (University of California San Diego), Sunny Dasgupta (Amazon)

🎯 What it does: Proposed the RMIR (Reasoning-Intensive Multimodal Image Retrieval) benchmark dataset to evaluate the performance of multimodal retrieval models on three categories of reasoning tasks: functional, temporal, and causal.

RNN as Linear Transformer: A Closer Investigation into Representational Potentials of Visual Mamba Models

Timing Yang (Johns Hopkins University), Guoyizhe Wei (Johns Hopkins University)

🎯 What it does: Systematically study the representational capability of Mamba in visual tasks and propose the Binary-AUC evaluation metric.

RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction

Chenxu Peng (Nankai University), Xiang Li (NKIARI)

🎯 What it does: Proposed a global-scale road segmentation dataset named WorldRoadSeg-360K (366,947 satellite images) and developed an interactive road extraction framework called RoadGIE based on connectivity-aware prompts.

Robo-SGG: Exploiting Layout-Oriented Normalization and Restitution Can Improve Robust Scene Graph Generation

Changsheng Lv (Beijing University of Posts and Telecommunications), Mengshi Qi (Beijing University of Posts and Telecommunications)

🎯 What it does: Proposed a pluggable module Robo-SGG that enhances the robustness of scene graph generation using layout information;

RobotSeg: A Model and Dataset for Segmenting Robots in Image and Video

Haiyang Mei (National University of Singapore), Mike Zheng Shou (National University of Singapore)

🎯 What it does: Propose RobotSeg, a robot segmentation model based on SAM 2 and the VRS video robot segmentation dataset, supporting image and video segmentation, and enabling zero-shot, single-point, box selection, and interactive refinement;

Robust Remote Sensing Image-Text Retrieval with Noisy Correspondence

Qiya Song (Hunan Normal University), Xudong Kang (Hunan University)

🎯 What it does: Investigate the noise correspondence problem in remote sensing image-text retrieval and propose a robust remote sensing image-text retrieval (RRSITR) framework.

Robust Spiking Neural Networks by Temporal Mutual Information

Mengting Xu (Zhejiang University), Gang Pan (Zhejiang University)

🎯 What it does: This paper proposes a regularization method based on Temporal Mutual Information (TMI) to enhance the robustness of Spiking Neural Networks (SNN).

Robustness Under Data Scarcity: Few-Shot Continual Adversarial Training for Evolving Threats

Wenxuan Wang (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)

🎯 What it does: Proposes a continuous adversarial training framework FS-CAT under few adversarial samples, and designs three key mechanisms: adversarial boundary loss, GMM prototype replay, and multi-domain balanced loss.

Role-SynthCLIP: A Role-Play Driven Diverse Synthetic Data Approach

Yuanxiang Huangfu (PatSnap Company Limited), Weilei Wang (PatSnap Company Limited)

🎯 What it does: Propose the Role-SynthCLIP framework, which enables multimodal large language models (MLLM) to generate multiple perspectives and fine-grained captions for the same image through multi-role prompting, thereby constructing high-quality, semantically rich synthetic image-text pairs.

Roots Beneath the Cut: Uncovering the Risk of Concept Revival in Pruning-Based Unlearning for Diffusion Models

Ci Zhang (University of Georgia), Geng Yuan (Stevens Institute of Technology)

🎯 What it does: Studied the safety of pruning in diffusion models and proposed a data- and training-agnostic attack framework. The framework uses low-rank matrix completion, Top-K sign preservation, and neuron maximum scaling to recover pruned weights and restore forgotten concepts; meanwhile, Gaussian blur defense is proposed.

RoSAMDepth: Robust Self-supervised Depth Estimation Leveraging Segment Anything Model

Xuanang Gao (Shanghai Jiao Tong University), Wei Liu (Shanghai Jiao Tong University)

🎯 What it does: This paper proposes RoSAMDepth, a framework that leverages object-level priors from the Segment Anything Model (SAM) to enhance the robustness of self-supervised monocular depth estimation. It injects SAM's segmentation information at the feature level and improves depth smoothing constraints and pseudo-label weighting methods, resulting in sharper and more accurate depth maps under various adverse conditions.

Rotation Invariant and Symmetry Aware Pixel Difference Network for Remote Sensing Object Detection

Jialei Zhan (National University of Defense Technology), Ming-Ming Cheng (Nankai University)

🎯 What it does: Proposed a remote sensing object detection network named RIS-PiDiNet that integrates rotation invariance and symmetry modeling.

Routing on Demand: DSNet for Efficient Progressive Point Cloud Denoising

Xiaoqian Cheng (University of Science and Technology of China), Renjie Chen (University of Science and Technology of China)

🎯 What it does: This paper proposes DSNet, a progressive denoising network for point clouds that achieves adaptive processing of local noise through a dynamic routing mechanism.

rPPG-VQA: A Video Quality Assessment Framework for Unsupervised rPPG Training

Tianyang Dai (University of Science and Technology of China), Yang Hu (University of Science and Technology of China)

🎯 What it does: Propose a video quality assessment framework specifically designed for unsupervised remote photoplethysmography (rPPG) training, named rPPG-VQA, which combines signal-level consensus SNR with scene disturbance evaluation based on a multi-modal large language model, and implements two-stage adaptive sampling based on quality scores.

RS-SSM: Refining Forgotten Specifics in State Space Model for Video Semantic Segmentation

Kai Zhu (Peking University), Jiahuan Zhou (University of Chinese Academy of Sciences)

🎯 What it does: Designed and implemented a video semantic segmentation framework RS-SSM based on the state space model (SSM), achieving more accurate pixel-level segmentation by compensating for the spatiotemporal details lost during SSM compression.

RunawayEvil: Jailbreaking the Image-to-Video Generative Models

Songping Wang (Nanjing University), Caifeng Shan (Nanjing University)

🎯 What it does: Propose a self-evolving multimodal jailbreak framework RunawayEvil based on the Strategy-Tactic-Action (S-T-A) paradigm, which can collaborate with text and image inputs to bypass the security mechanisms of Image-to-Video (I2V) models.

RxnCaption: Reformulating Reaction Diagram Parsing as Visual Prompt Guided Captioning

Jiahe Song (Shanghai Jiao Tong University), Conghui He (Shanghai Artificial Intelligence Laboratory)

🎯 What it does: Redefined the chemical reaction diagram parsing task as an image description task, proposed and implemented the RxnCaption framework;

S$^2$-MLLM: Boosting Spatial Reasoning Capability of MLLMs for 3D Visual Grounding with Structural Guidance

Beining Xu, Hesheng Wang (Shanghai Jiao Tong University)

🎯 What it does: Propose the S-MLLM framework, which equips multi-modal large language models with implicit 3D spatial reasoning capabilities by incorporating feed-forward 3D reconstruction as spatial guidance during training, enabling 3D visual localization tasks.

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

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

🎯 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;

SAGA: Source Attribution of Generative AI Videos

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

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

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

🎯 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.

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

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

🎯 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.

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)

🎯 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).

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

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

🎯 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;

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

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

🎯 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.

SASNet: Spatially-Adaptive Sinusoidal Networks for INRs

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

🎯 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).

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

Ruixiang Zhao, Xirong Li (Renmin University Of China)

🎯 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.

Scalable Multi-View Subspace Clustering with Tensorized Anchor Guidance

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

🎯 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)

🎯 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.

Scaling Agentic Reinforcement Learning for Tool-Integrated Reasoning in VLMs

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

🎯 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)

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

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

🎯 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)

🎯 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.

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

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

🎯 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.

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)

🎯 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;

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

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

🎯 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.

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

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

🎯 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.

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))

🎯 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.

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)

🎯 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)

🎯 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.

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

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

🎯 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.

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

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

🎯 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)

🎯 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)

🎯 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.

See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs

Yongchang Zhang (Southeast University), Yang Chen (Southeast University)

🎯 What it does: Propose a no-training, plug-and-play visual evidence-driven decoding framework ECRD, which real-time supervises the multimodal reasoning process to avoid visual hallucinations.

See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning

Shuoshuo Zhang (Tsinghua University), Rui Wang (Microsoft Research Asia)

🎯 What it does: Propose the BiPS method, which enhances the perception capability of VLMs for fine-grained visual evidence by employing evidence-preserving and evidence-ablated views under problem conditions during training, through bidirectional KL constraints.