CVPR 2025 Papers — Page 24

Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection

Ji Du (Nankai University), Ping Li (Hong Kong Polytechnic University)

🎯 What it does: An unsupervised covert object detection framework EASE is proposed, which achieves object segmentation by first identifying the environment and then inferring the target.

ShiftwiseConv: Small Convolutional Kernel with Large Kernel Effect

Dachong Li (Shenzhen University), Jianqiang Li

🎯 What it does: This paper proposes the Shiftwise (SW) convolution module, which combines 3×3 standard convolution with the Shift operation to achieve a receptive field and long-range dependencies comparable to large convolutions, thereby replacing large convolution kernels in a pure CNN framework.

Shining Yourself: High-Fidelity Ornaments Virtual Try-on with Diffusion Model

Yingmao Miao (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)

🎯 What it does: This paper proposes the first virtual try-on task for accessories and designs a pose-aware mask prediction and mask-guided attention mechanism to achieve high-fidelity and geometrically consistent try-on effects.

ShotAdapter: Text-to-Multi-Shot Video Generation with Diffusion Models

Ozgur Kara (University of Illinois Urbana-Champaign), Tobias Hinz (Adobe)

🎯 What it does: The ShotAdapter framework is proposed, which fine-tunes a single-shot text-to-video model into a multi-shot video generator, allowing for independent control over the duration, content, and background of each shot.

Show and Segment: Universal Medical Image Segmentation via In-Context Learning

Yunhe Gao (Microsoft), Dimitris N. Metaxas (Rutgers University)

🎯 What it does: The Iris framework is proposed, which achieves unsupervised adaptation segmentation of 3D medical images using a single reference image-label pair, supporting new tasks and new anatomical structures without the need for model fine-tuning.

Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models

Itay Benou (Ben-Gurion University of the Negev), Tammy Riklin Raviv (Ben-Gurion University of the Negev)

🎯 What it does: Transform any visual neural network into an interpretable spatially aware concept bottleneck model, outputting both concept explanations and corresponding spatial heatmaps during inference, supporting interactive exploration and local parameter tuning.

ShowHowTo: Generating Scene-Conditioned Step-by-Step Visual Instructions

Tomáš Souček (Czech Technical University in Prague), Josef Sivic (Czech Technical University in Prague)

🎯 What it does: This paper proposes a system based on video diffusion models that can automatically generate a visually guided sequence that is consistent with the scene and unfolds step by step, given a starting scene image and text instructions.

ShowMak3r: Compositional TV Show Reconstruction

Sangmin Kim (Seoul National University), Jaesik Park (Seoul National University)

🎯 What it does: This paper presents ShowMak3r, a pipeline for constructing editable dynamic radiance fields based on single-camera television program videos.

ShowUI: One Vision-Language-Action Model for GUI Visual Agent

Kevin Qinghong Lin (National University of Singapore), Mike Zheng Shou (National University of Singapore)

🎯 What it does: An end-to-end visual-language-action model called ShowUI is proposed and trained for executing localization and navigation tasks in GUI environments.

SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model

Zhenglin Huang (University of Liverpool), Guangliang Cheng (University of Liverpool)

Silence is Golden: Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation

Yuan Gan (Zhejiang University), Yi Yang (Zhejiang University)

🎯 What it does: A two-stage active protection method called Silencer is proposed to generate robust adversarial perturbations in audio-driven speaker animation models based on LDM, thereby protecting portrait privacy.

Silent Branding Attack: Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models

Sangwon Jang (Korea Advanced Institute of Science and Technology), Sung Ju Hwang (Korea Advanced Institute of Science and Technology)

🎯 What it does: This paper proposes the Silent Branding Attack, a silent brand placement attack that does not require text triggers. It covertly embeds brand logos in the training data, allowing the text-to-image diffusion model to naturally generate the specified logo during creation.

SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation

Leigang Qu (National University of Singapore), Tat-Seng Chua (National University of Singapore)

🎯 What it does: A self-improvement framework SILMM is proposed to enhance the text-image alignment in complex text-to-image generation for multimodal models.

Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations

Ahmad Rahimi (Ecole Polytechnique Federale de Lausanne), Alexandre Alahi (Ecole Polytechnique Federale de Lausanne)

🎯 What it does: This paper studies causal representation in multi-agent interactions, proposing a regularization method to enhance the model's perception of causal relationships and achieve causal transfer from simulation to real scenarios.

SimAvatar: Simulation-Ready Avatars with Layered Hair and Clothing

Xueting Li (NVIDIA), Umar Iqbal (NVIDIA)

🎯 What it does: Generate 3D clothing, hairstyles, and layered human avatars that can be directly used for physical or neural simulations from text prompts.

Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking

Chaocan Xue (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)

🎯 What it does: This study investigates the layer redundancy issue in lightweight ViT trackers and proposes a similarity-guided layer adaptive mechanism that selects a single representative layer to prune redundant layers, achieving real-time drone target tracking.

SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment

Katrin Renz (Wayve), Oleg Sinavski (Wayve)

🎯 What it does: Proposes the SimLingo model, integrating closed-loop driving, visual language understanding, and language-action alignment tasks, achieving LiDAR-free autonomous driving based on cameras.

SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection

Phi Vu Tran (LexisNexis Risk Solutions)

🎯 What it does: This paper proposes the SimLTD framework, which achieves long-tail object detection through staged training (pre-training head classes, transferring tail classes, and fine-tuning sample sets) and optionally uses unlabeled images for semi-supervised training.

SimMotionEdit: Text-Based Human Motion Editing with Motion Similarity Prediction

Zhengyuan Li (Purdue University), Aniket Bera (Purdue University)

🎯 What it does: This paper proposes SimMotionEdit, a text-based 3D human motion editing framework that incorporates a motion similarity prediction auxiliary task into the editing task through multi-task learning.

Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion

Emiel Hoogeboom (Google Deepmind), Tim Salimans (Google Deepmind)

🎯 What it does: This paper presents Simpler Diffusion (SiD2), an end-to-end diffusion model based on pixel space, which achieves 1.5 FID and reaches state-of-the-art (SOTA) visual quality on the ImageNet datasets of 512, 128, 256, 1024, and the Kinetics600 video dataset.

Simplification Is All You Need against Out-of-Distribution Overconfidence

Keke Tang (Guangzhou University), Zhihong Tian (Guangzhou University)

🎯 What it does: The paper proposes to alleviate the overconfidence of deep neural networks on OOD samples through model simplification (knowledge distillation to reduce capacity, selective removal of ReLU to reduce non-linearity).

Simulator HC: Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision

Xinyue Zhang (ShanghaiTech University), Laurent Kneip (ShanghaiTech University)

🎯 What it does: This paper proposes a three-stage solving process: first, a one-dimensional convolutional regression network predicts the initial solution for geometric problems; then, an online simulator generates corresponding consistent problem-solution pairs based on that prediction; finally, the Homotopy Continuation (HC) method with single root tracking is used to obtain precise solutions from the initial pair, successfully applied to generalized camera relocalization and generalized relative pose scaling problems.

SimVS: Simulating World Inconsistencies for Robust View Synthesis

Alex Trevithick (University of California San Diego), Pratul P. Srinivasan (Google DeepMind)

🎯 What it does: This paper proposes the use of video diffusion models to generate motion and lighting inconsistencies, training a multi-view harmonization network to recover consistent 3D scenes from sparse and inconsistent captures.

Single Domain Generalization for Few-Shot Counting via Universal Representation Matching

Xianing Chen (Huawei Noah's Ark Lab), Xinghao Chen (Huawei Noah's Ark Lab)

🎯 What it does: A few-shot counting model for single-source domain generalization, URM, is proposed, which achieves cross-domain counting through distilled universal visual-language prototypes from CLIP.

SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors

Yufan Wu (Shanghai Jiao Tong University), Wenjun Zhang (Shanghai Jiao Tong University)

🎯 What it does: Generating animatable high-quality 3D human avatars from single portrait images

SINR: Sparsity Driven Compressed Implicit Neural Representations

Dhananjaya Jayasundara (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)

🎯 What it does: Developed the SINR algorithm, which utilizes high-dimensional sparse codes in the INR weight space for compression, compatible with any INR-based compression scheme.

SIR-DIFF: Sparse Image Sets Restoration with Multi-View Diffusion Model

Yucheng Mao (University of Michigan), Jeong Joon Park (University of Michigan)

🎯 What it does: A multi-view image recovery framework SIR-Diff is proposed, utilizing a multi-view diffusion model to jointly denoise blurred or low-resolution images while maintaining 3D consistency.

Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models

Jie Ren (Michigan State University), Lingjuan Lyu (Sony AI)

🎯 What it does: This paper proposes a benchmark for concept removal in text-to-image diffusion models, including a dataset of six categories of undesirable concepts called Six-CD and a Dual-Version Dataset, and evaluates the performance of ten removal methods.

SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons

Yuanyou Xu (Zhejiang University), Yi Yang (Harvard University)

🎯 What it does: A multi-view and 3D generation system called SKDream has been developed, which can be used for arbitrary skeletal conditions, enabling instant generation from skeletons to multi-view images and then to textured 3D models.

SKE-Layout: Spatial Knowledge Enhanced Layout Generation with LLMs

Junsheng Wang (Chongqing University), Chao Chen (Chongqing University)

🎯 What it does: A framework named SKE-Layout is proposed, which generates precise 2D/3D layouts from text descriptions by injecting multi-source spatial knowledge into large language models.

Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch

Aneeshan Sain (University of Surrey), Yi-Zhe Song (University of Surrey)

🎯 What it does: This paper proposes an efficient retrieval model for human-drawn sketches.

SketchAgent: Language-Driven Sequential Sketch Generation

Yael Vinker (Massachusetts Institute of Technology), Antonio Torralba (Massachusetts Institute of Technology)

🎯 What it does: Developed SketchAgent, a language-driven, sequential sketch generation system based on multimodal LLM, which supports real-time drawing, editing, and collaborative sketching through natural language and dialogue interaction.

SketchFusion: Learning Universal Sketch Features through Fusing Foundation Models

Subhadeep Koley (University of Surrey), Yi-Zhe Song (University of Surrey)

🎯 What it does: This paper proposes a hybrid feature extraction framework that integrates Stable Diffusion and CLIP to generate general sketch features.

Sketchtopia: A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback

Mohd Hozaifa Khan (Indian Institute of Information Technology Hyderabad), Ravi Kiran Sarvadevabhatla (Indian Institute of Information Technology Hyderabad)

🎯 What it does: This paper presents the Sketchtopia dataset and an AI framework based on asynchronous multimodal communication, which includes two types of agents: the drawer (DRAWBOT) and the guesser (GUESSBOT), and implements a human-machine hybrid Pictionary gameplay.

SketchVideo: Sketch-based Video Generation and Editing

Feng-Lin Liu (Chinese Academy of Sciences), Lin Gao (Chinese Academy of Sciences)

🎯 What it does: We propose SketchVideo, a sketch-based controllable video generation and editing framework that supports sketching on keyframes and generating multi-frame videos combined with text prompts, as well as performing local edits on real videos.

Sketchy Bounding-box Supervision for 3D Instance Segmentation

Qian Deng (Nankai University), Jian Yang (Nankai University)

🎯 What it does: A Sketchy-3DIS framework is proposed to achieve weakly supervised 3D instance segmentation with only imprecise sketchy bounding boxes provided.

SkillMimic: Learning Basketball Interaction Skills from Demonstrations

Yinhuai Wang (Hong Kong University of Science and Technology), Ping Tan (Hong Kong University of Science and Technology)

🎯 What it does: Developed the SkillMimic framework, enabling humanoid robots in physical simulations to master various interaction skills by learning from basketball demonstrations in the human-object interaction (HOI) dataset, which can be combined by a high-level controller to accomplish complex tasks such as scoring consecutively.

Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves

Shihan Wu (University of Electronic Science and Technology of China), Heng Tao Shen (Tongji University)

🎯 What it does: This paper proposes the Skip Tuning method, which performs layer skipping (LSkip) and category skipping (CSkip) directly on the CLIP pre-trained model to achieve transfer learning for downstream tasks without adding extra prompt vectors or adapters.

SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling

Qi Zhu (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)

🎯 What it does: This paper proposes an open-world interpretation framework for remote sensing images, SkySense-O, which includes a newly created fine-grained annotated dataset, Sky-SA, as well as a vision-centered visual-language alignment and training strategy.

SLADE: Shielding against Dual Exploits in Large Vision-Language Models

Md Zarif Hossain (Southern Illinois University Carbondale), Ahmed Imteaj (Southern Illinois University Carbondale)

🎯 What it does: Unsupervised adversarial fine-tuning of the CLIP visual encoder in large visual-language models is performed, using dual-layer contrastive learning (patch-level and global-level) to enhance robustness against gradient attacks and optimization-based jailbreak attacks.

SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos

Yuzheng Liu (Peking University), Baoquan Chen (Peking University)

🎯 What it does: SLAM3R is proposed, a system that achieves real-time, high-definition dense 3D reconstruction using RGB video, employing a two-layer neural network framework for end-to-end point cloud prediction and global registration;

SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models

Zilan Wang (Nanyang Technological University), Zhengzhong Tu (University of Maryland)

🎯 What it does: A watermark embedding method for text-to-image diffusion models is proposed, which can maintain recognizable ownership identifiers even after the model is fine-tuned downstream.

SlideChat: A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding

Ying Chen (Shanghai AI Laboratory), Junjun He (Shanghai AI Laboratory)

🎯 What it does: Developed and released SlideChat, an assistant capable of understanding and engaging in multimodal dialogue on whole slide images, and constructed the corresponding large-scale instruction dataset SlideInstruction and evaluation benchmark SlideBench.

SLVR: Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation

Ning Ni (Beijing Normal University), Libao Zhang (Beijing Normal University)

🎯 What it does: A super lightweight visual reconstruction framework SLVR has been developed, introducing two new modules: B2Conv and FDEL.

SmartCLIP: Modular Vision-language Alignment with Identification Guarantees

Shaoan Xie (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)

🎯 What it does: Proposes SmartCLIP, which achieves visual-text alignment through adaptive masking, addressing the issues of information mismatch and representation mixing in CLIP.

SmartEraser: Remove Anything from Images using Masked-Region Guidance

Longtao Jiang (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)

🎯 What it does: A method for object removal based on Masked-Region Guidance called SmartEraser is proposed, and a large-scale Syn4Removal dataset with 1M triplets is constructed.

SMILE: Infusing Spatial and Motion Semantics in Masked Video Learning

Fida Mohammad Thoker (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)

🎯 What it does: Proposes the SMILE framework, which uses high-level semantic features from the CLIP pre-trained model and synthetic motion data for self-supervised learning in a video mask autoencoder.

SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity

Yijie Xu (Hangzhou Dianzi University), Chenggang Yan (Hangzhou Dianzi University)

🎯 What it does: A multi-language, multi-modal, and temporal social media popularity prediction benchmark (SMTPD) has been constructed, and a baseline model based on multi-modal feature fusion and LSTM regression has been proposed to predict the popularity time series of YouTube videos within 30 days.

SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device

Yushu Wu (Snap Inc.), Jian Ren (Snap Inc.)

🎯 What it does: A complete acceleration framework is proposed for real-time inference of 5-second text-to-video generation on mobile devices.

SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training

Jierun Chen (Snap Inc.), Jian Ren (Snap Inc.)

🎯 What it does: A text-to-image diffusion model called SnapGen is proposed, which minimizes and can generate 1024×1024 high-resolution images in real-time on mobile devices.

SnowMaster: Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization

Jianyu Lai (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)

🎯 What it does: SnowMaster proposes a single-image snow removal method based on a multimodal large language model and a multi-angle evaluation framework, enhancing the effectiveness of real-world snow scene removal through semi-supervised training.

SOAP: Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection

Hyo-Jun Lee (Kangwon National University), Jinu Lee (42dot Inc.)

🎯 What it does: A visual-driven 3D semantic scene completion method named SOAP is proposed, which combines occlusion-aware view projection and a scene-adaptive decoder;

SocialGesture: Delving into Multi-person Gesture Understanding

Xu Cao (University of Illinois Urbana-Champaign), James M. Rehg (University of Illinois Urbana-Champaign)

🎯 What it does: This paper presents and releases the SocialGesture dataset, focusing on the understanding of natural social gestures among multiple people, and based on this, designs three benchmark tasks: gesture recognition, temporal localization, and visual question answering.

SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction

Kai Chen (Nanjing University of Aeronautics and Astronautics), Ziyuan Wang (Beihang University)

🎯 What it does: The SocialMOIF model is proposed, which combines multi-order intent fusion, trajectory distribution approximator, and global trajectory optimizer to achieve multi-modal and parallel trajectory prediction for pedestrians (or vehicles, athletes).

Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation

Zhongwen Zhang (University of Waterloo), Yuri Boykov (University of Waterloo)

🎯 What it does: This paper proposes a soft self-labeling framework that utilizes the second-order relaxation of the Potts model and jointly optimizes the network with soft pseudo-labels, achieving end-to-end training in the scribble weakly supervised semantic segmentation task.

SoftShadow: Leveraging Soft Masks for Penumbra-Aware Shadow Removal

Xinrui Wang (Nanyang Technological University), Bihan Wen (Nanyang Technological University)

🎯 What it does: The SoftShadow framework is proposed, which generates soft shadow masks and completes shadow removal without the need for pre-defined shadow masks.

SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer

Hao Chen (Carnegie Mellon University), Emad Barsoum (AMD)

🎯 What it does: This paper proposes SoftVQ-VAE, a differentiable continuous image tokenizer that significantly reduces the number of tokens to 32 or 64 through soft category posterior aggregation of multiple codewords, while maintaining high-quality reconstruction and improving the efficiency of generative models.

SOGS: Second-Order Anchor for Advanced 3D Gaussian Splatting

Jiahui Zhang (Nanyang Technological University), Shijian Lu (Nanyang Technological University)

🎯 What it does: A 3D Gaussian scattering method based on second-order anchor points (SOGS) is proposed, which can reduce the dimensionality of anchor point features while maintaining or even improving rendering quality.

SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters

Jianping Jiang (SenseTime Research), Ziwei Liu (Nanyang Technological University)

🎯 What it does: We propose and implement SOLAMI, an end-to-end social visual-language-action (VLA) framework that enables 3D autonomous characters to interact with users in an immersive manner through voice and body language in VR environments.

SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving

Xuesong Chen (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)

🎯 What it does: The SOLVE framework is proposed, which achieves collaboration between large visual language models (VLM) and end-to-end (E2E) networks at the feature layer and trajectory layer, addressing the complexity and interpretability issues of driving decisions.

Solving Instance Detection from an Open-World Perspective

Qianqian Shen (Zhejiang University), Shu Kong (University of Macau)

🎯 What it does: The IDOW method is proposed, which significantly improves instance detection performance by performing metric learning on the base model using open-world data, combined with interference sample sampling and NeRF new perspective synthesis.

SoMA: Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning

Seokju Yun (University of Seoul), Youngmin Ro (University of Seoul)

🎯 What it does: This paper proposes a parameter-efficient fine-tuning framework called SoMA based on singular value decomposition, aimed at adapting to domain transfer tasks while maintaining the generalization capability of visual foundation models.

Sonata: Self-Supervised Learning of Reliable Point Representations

Xiaoyang Wu (University of Hong Kong), Julian Straub (Meta Reality Labs Research)

🎯 What it does: We propose Sonata, a self-supervised learning framework for point clouds that achieves strong linear probing performance with very few trainable parameters and supports multi-scale representations and optional decoders.

Sonic: Shifting Focus to Global Audio Perception in Portrait Animation

Xiaozhong Ji (Tencent), Chengjie Wang (Tencent)

🎯 What it does: The Sonic framework is proposed, utilizing global audio perception to achieve precise lip synchronization, rich expressions, and head movements in facial animation without the need for visual motion guidance.

Sound Bridge: Associating Egocentric and Exocentric Videos via Audio Cues

Sihong Huang (South China University of Technology), Yaowei Wang (Harbin Institute of Technology)

🎯 What it does: This study proposes using audio as a bridge to align first-person and third-person videos from different perspectives, and learns the audio perception representation space through audio-visual and audio-text cross-attention modules.

SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding

Mingfei Chen (University of Washington), Alexander Richard

🎯 What it does: SoundVista is proposed, which generates realistic binaural environmental sounds from sparse reference microphone recordings and visual data at any new viewpoint;

SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts

Shijia Zhao (Xiamen University), Cheng Wang (Xiamen University)

🎯 What it does: A two-stage sparse supervised 3D object detection enhancement framework SP3D is proposed, which utilizes multimodal models to generate cross-modal semantic hints for producing high-quality pseudo-labels, pre-training the detector first and then fine-tuning it;

SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models

Yongting Zhang (University of Science and Technology of China), Jing Shao (Shanghai Artificial Intelligence Laboratory)

🎯 What it does: This paper proposes and constructs the SPA-VL secure preference alignment dataset for training the safety of visual-language models (VLMs); it aligns models such as LLaVA using RLHF methods (PPO, DPO) and evaluates them on multiple safety and generality benchmarks.

SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images

Zixuan Huang (Stability AI), Varun Jampani (Stability AI)

🎯 What it does: A two-stage single-view 3D reconstruction method called SPAR3D is proposed, which first generates a sparse point cloud using point diffusion, and then combines the point cloud with images using a tri-plane Transformer to generate high-quality textured meshes.

SPARC: Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models

Kevin Miller (Boston University), Venkatesh Saligrama (Boston University)

🎯 What it does: This paper proposes a zero-shot multi-label recognition framework called SPARC, based on visual-language model (VLM) black-box scoring, which utilizes normalization to eliminate image and prompt biases, and enhances classification performance through composite prompts and adaptive rank fusion.

SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction

Yutao Tang (Johns Hopkins University), Cheng Peng (Johns Hopkins University)

🎯 What it does: To achieve high-quality novel view synthesis in sparse view scenarios, the SPARS3R method is proposed. It first aligns dense point clouds to the structure-from-motion (SfM) sparse point clouds, and then uses semantic segmentation for local alignment of erroneous points, resulting in accurate poses and detail-rich 3D point clouds.

Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians

Changfeng Ma (Nanjing University), Yanwen Guo (Nanjing University)

🎯 What it does: A method is proposed to directly predict 2D Gaussians from point clouds and perform lighting rendering, capable of generating high-quality images directly from sparse point clouds.

Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering

Cheng Sun (NVIDIA), Yu-Chiang Frank Wang (National Taiwan University)

🎯 What it does: We propose SVRaster, a differentiable radiance field rendering framework based on sparse voxels, which achieves real-time high-quality view synthesis through rasterization.

Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views

Jiang Wu (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)

🎯 What it does: A sparse view 3D surface reconstruction method based on Gaussian partitioning, called Sparse2DGS, is proposed.

SparseAlign: a Fully Sparse Framework for Cooperative Object Detection

Yunshuang Yuan (Munich Center for Machine Learning), Monika Sester (Leibniz University Hannover)

🎯 What it does: A fully sparse framework called SparseAlign has been designed and implemented for LiDAR-based collaborative object detection, enhancing feature connectivity through sparse convolution to address the issues of missing central features and isolated convolutions.

Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis

Woojung Han (Yonsei University), Seong Jae Hwang (Yonsei University)

🎯 What it does: A training-free, attention map-based spatial transport optimization method (STORM) is proposed to achieve precise spatial alignment of objects in text-to-image generation.

Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation

Qi Lv (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)

🎯 What it does: A dual-arm robot manipulation strategy based on a spatio-temporal graph diffusion network is proposed and implemented. This strategy generates action sequences that avoid collisions and satisfy kinematic constraints by combining the robot's structure and joint movements.

Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models

Xingrui Wang (Johns Hopkins University), Alan Yuille (Johns Hopkins University)

🎯 What it does: Created the Spatial457 benchmark dataset, designed with 5 difficulty levels and 7 types of question-answer tasks, for the systematic evaluation of large multimodal models (LMM) in 6D spatial reasoning (multi-object recognition, 2D position, 3D position, 3D pose).

SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language

Zehan Wang (Zhejiang University), Zhou Zhao (Zhejiang University)

🎯 What it does: This paper proposes SpatialCLIP, which enhances CLIP's understanding of spatial relationships by improving the visual encoder and language supervision, and builds SpatialLLaVA based on this model to enhance the spatial intelligence of multimodal large models.

SpatialDreamer: Self-supervised Stereo Video Synthesis from Monocular Input

Zhen Lv (Alibaba Group), Dian Zheng (Sun Yat-sen University)

🎯 What it does: This paper proposes a self-supervised stereo video synthesis framework based on a video diffusion model—SpatialDreamer, which can generate high-quality stereo videos from monocular videos.

SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models

Wufei Ma (Johns Hopkins University), Jieneng Chen (Johns Hopkins University)

🎯 What it does: Proposes SpatialLLM, a multimodal large model capable of understanding and reasoning about three-dimensional spatial relationships;

Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting

Jingyi Xu (Shanghai Jiao Tong University), Ling Pei (Shanghai Jiao Tong University)

🎯 What it does: A spatial and temporal decoupled visual 3D vehicle occupancy prediction framework, EfficientOCF, is proposed, achieving efficient occupancy prediction within this framework.

Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling

Junha Hyung (KAIST), Jaegul Choo (KAIST)

🎯 What it does: A training-free sampling guidance method named Spatiotemporal Skip Guidance (STG) is proposed to enhance the generation quality of Transformer-based video diffusion models.

SPC-GS: Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs

Guibiao Liao (Peking University), Kanglin Liu (Pengcheng Laboratory)

🎯 What it does: A framework for free-viewpoint synthesis of sparse views in indoor open worlds based on 3D Gaussian Splatting, called SPC-GS, is proposed.

Spectral Informed Mamba for Robust Point Cloud Processing

Ali Bahri (École de technologie supérieure), Christian Desrosiers (École de technologie supérieure)

🎯 What it does: This paper introduces a sequence traversal method based on graph features, applying the Mamba network to point cloud classification, segmentation, and self-supervised pre-training, achieving more robust feature extraction.

Spectral State Space Model for Rotation-Invariant Visual Representation Learning

Sahar Dastani (École de Technologie Supérieure), Christian Desrosiers (École de Technologie Supérieure)

🎯 What it does: This paper presents Spectral VMamba, a visual state space model that constructs rotation-invariant paths using the graph Laplacian spectrum.

SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting

Jiajun Tang (Peking University), Boxin Shi (Peking University)

🎯 What it does: In the 3D Gaussian splatting framework, high-frequency reflections and indirect lighting of highly specular surfaces are simulated using the physical rendering equation and ray tracing, achieving editable rendering effects.

SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes

Cheng-De Fan (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)

🎯 What it does: Combining 3D Gaussian Splatting, physics-based rendering, and deformable fields to achieve the reconstruction and high-quality rendering of dynamic mirror scenes.

Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives

Alex Hanson (University of Maryland), Tom Goldstein (University of Maryland)

🎯 What it does: This paper proposes the Speedy-Splat method, which significantly improves the rendering speed, model compression, and training efficiency of 3D Gaussian Splatting by accurately locating Gaussian primitives and introducing soft and hard pruning.

SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception

Yaniv Benny (Tel Aviv University), Lior Wolf (Tel Aviv University)

🎯 What it does: This paper presents SphereUFormer, a U-shaped Transformer capable of directly operating on spherical 360° images for depth estimation and semantic segmentation.

Spherical Manifold Guided Diffusion Model for Panoramic Image Generation

Xiancheng Sun, Gang Shen

🎯 What it does: Proposed spherical manifold convolution (SMConv) and a diffusion model based on SMConv (SMGD) for generating panoramic images from text.

Spiking Transformer with Spatial-Temporal Attention

Donghyun Lee (Yale University), Priyadarshini Panda (Yale University)

🎯 What it does: A block-based space-time attention mechanism STAtten is proposed for pulse Transformers, integrating spatial and temporal information to enhance feature representation.

Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer

Yufei Guo (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Intelligent Science and Technology Academy of CASIC)

🎯 What it does: An accurate additive pulse self-attention (A2OSA) model is proposed, combining Transformer with spiking neural networks (SNN);

SpiritSight Agent: Advanced GUI Agent with One Look

Zhiyuan Huang (SenseTime Research), Mingjie Zhan (SenseTime Research)

🎯 What it does: Designed and implemented a vision-based end-to-end GUI agent called SpiritSight, capable of performing high-precision, low-latency GUI navigation tasks across multiple platforms.

Spk2SRImgNet: Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering

Yuanlin Wang (Peking University), Tiejun Huang (Harbin Institute of Technology)

🎯 What it does: This paper proposes the Spk2SRImgNet network, which utilizes motion-aligned collaborative filtering to achieve super-resolution reconstruction from low-resolution pulsed streams to high-resolution images.

SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving

Georg Hess (Zenseact), Lennart Svensson (Chalmers University of Technology)

🎯 What it does: This paper presents SplatAD, which achieves real-time high-quality synthesis of camera and LiDAR data based on 3D Gaussian splatting rendering.

SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis

Hyojun Go (EverEx), Changick Kim (KAIST)

🎯 What it does: Proposes the SplatFlow framework, which can directly generate and edit 3D Gaussian Splatting scenes from text, achieving multi-view generation and editing with a single training session.

SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving

Su Sun (Purdue University), Mei Chen (Microsoft)

🎯 What it does: A self-supervised dynamic Gaussian scattering method called SplatFlow is designed, which combines Neural Motion Flow Field to achieve 4D spatiotemporal representation without 3D bounding box supervision, and is used for dynamic scene reconstruction and viewpoint synthesis in autonomous driving scenarios.

Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images

Zheng Chen (Tsinghua University), Song-Hai Zhang

🎯 What it does: Designed and implemented Splatter-360, an end-to-end 3D Gaussian light scattering framework for real-time synthesis of new views from wide baseline panoramic images.

SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video

Jongmin Park (Korea Advanced Institute of Science and Technology), Munchurl Kim (Korea Advanced Institute of Science and Technology)

🎯 What it does: Developed a real-time 3D Gaussian model that can synthesize high-quality dynamic views from monocular video without COLMAP;