CVPR 2025 Papers — Page 23

Scaling Down Text Encoders of Text-to-Image Diffusion Models

Lifu Wang (JD.com Inc.), Xiaodong He (Georgia Institute of Technology)

🎯 What it does: The T5-XXL model is scaled down to a smaller model through visual knowledge distillation, maintaining nearly the same generation quality in text-to-image diffusion models.

Scaling Inference Time Compute for Diffusion Models

Nanye Ma (New York University), Saining Xie (Google)

🎯 What it does: This paper proposes a framework to enhance generation quality during the inference phase of diffusion models by searching for noise and combining various validators and search algorithms.

Scaling Mesh Generation via Compressive Tokenization

Haohan Weng (South China University of Technology), C.L. Philip Chen

🎯 What it does: A block and patch compression tokenization (BPT) method is proposed, which reduces the length of mesh sequences by approximately 75%, thereby supporting the generation of over 8k face meshes.

Scaling Properties of Diffusion Models For Perceptual Tasks

Rahul Ravishankar (University of California Berkeley), Jitendra Malik (University of California Berkeley)

🎯 What it does: A unified modeling of perception tasks such as depth estimation, optical flow estimation, and instance segmentation is performed using an iterative inference framework based on diffusion models, and performance is enhanced through scaling strategies during the pre-training, fine-tuning, and inference phases.

Scaling up Image Segmentation across Data and Tasks

Pei Wang (Amazon), Stefano Soatto (Amazon)

🎯 What it does: A scalable image segmentation framework called MQ-Former is proposed, capable of joint training on various datasets and tasks, achieving open vocabulary and free-form segmentation.

Scaling Vision Pre-Training to 4K Resolution

Baifeng Shi (NVIDIA), Hongxu Yin (NVIDIA)

🎯 What it does: The PS3 framework extends CLIP-style visual pre-training to 4K resolution and utilizes it as a visual encoder to build a high-resolution multimodal large language model VILA-HD, which is further evaluated on a self-constructed 4KPro benchmark.

ScaMo: Exploring the Scaling Law in Autoregressive Motion Generation Model

Shunlin Lu (Sun Yat-sen University), Ruimao Zhang (Sun Yat-sen University)

🎯 What it does: A scalable text-driven motion generation framework ScaMo is proposed, combining FSQ-VAE and prefix autoregressive Transformer, and systematically studying the scaling laws of motion generation tasks.

SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting

Chenyu Zhang (Peking University), Jiahuan Zhou (Peking University)

🎯 What it does: To address the issue of adaptive testing of visual-language models under domain shift conditions, a Supportive Clique Attribute Prompt (SCAP) framework is proposed, enabling cross-sample information fusion and adaptation during transductive testing.

Scenario Dreamer: Vectorized Latent Diffusion for Generating Driving Simulation Environments

Luke Rowe (Mila), Felix Heide (Princeton University)

🎯 What it does: This paper presents Scenario Dreamer, a fully data-driven generative simulator for autonomous driving planning.

Scene Map-based Prompt Tuning for Navigation Instruction Generation

Sheng Fan (Zhejiang University), Yi Yang (Zhejiang University)

🎯 What it does: A navigation instruction generation framework called MAPINSTRUCTOR is proposed, which is based on scene map prompt tuning. It combines 3D voxel scene encoding, global topological map prompts, and landmark uncertainty assessment to achieve more accurate navigation instruction generation.

Scene Splatter: Momentum 3D Scene Generation from Single Image with Video Diffusion Model

Shengjun Zhang (Tsinghua University), Yueqi Duan (Tsinghua University)

🎯 What it does: A momentum-based video diffusion framework is designed to iteratively generate high-fidelity and consistent 3D Gaussian scenes from a single image.

Scene-agnostic Pose Regression for Visual Localization

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

🎯 What it does: This paper proposes the Scene-Independent Pose Regression (SPR) task, designs a dual-branch SPR-Mamba model, and constructs the 360SPR dataset with 200K panoramic images to address issues such as the need for retraining in absolute pose regression, reliance on retrieval databases in relative pose regression, and cumulative drift in visual odometry.

Scene-Centric Unsupervised Panoptic Segmentation

Oliver Hahn, Stefan Roth

🎯 What it does: CUPS is proposed, an unsupervised panoramic segmentation framework based on scene centers, which generates high-resolution pseudo-labels using depth, motion, and visual features, and enhances model performance through self-supervised training.

Scene4U: Hierarchical Layered 3D Scene Reconstruction from Single Panoramic Image for Your Immerse Exploration

Zilong Huang (Sun Yat-sen University), Ting Han (Sun Yat-sen University)

🎯 What it does: Based on a single panoramic image, a hierarchical reconstruction framework called Scene4U is proposed, capable of generating high-quality, dynamic obstacle-free, and freely explorable 3D scenes.

SceneCrafter: Controllable Multi-View Driving Scene Editing

Zehao Zhu (University of Texas at Austin), Dragomir Anguelov (Waymo)

🎯 What it does: We propose SceneCrafter, a model for editing driving scenes that achieves realistic and 3D-consistent results from multiple camera perspectives, supporting local foreground object insertion/deletion as well as global weather/time editing.

SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model

Shuhan Tan (UT Austin), Chiyu Max Jiang (Waymo LLC)

🎯 What it does: Designed and implemented SceneDiffuser++, a unified diffusion-based world model capable of generating complete scenes from maps and start-end points at a city scale, performing behavior prediction, occlusion reasoning, dynamic agent generation (creation/deletion), and traffic light state simulation, achieving end-to-end travel-level traffic simulation.

SceneFactor: Factored Latent 3D Diffusion for Controllable 3D Scene Generation

Aleksey Bokhovkin, Angela Dai (Technical University of Munich)

🎯 What it does: This paper presents SceneFactor, a controllable and editable 3D scene generation method implemented through a hierarchical latent diffusion model.

SceneTAP: Scene-Coherent Typographic Adversarial Planner against Vision-Language Models in Real-World Environments

Yue Cao (Agency for Science Technology and Research), Qing Guo (Agency for Science Technology and Research)

🎯 What it does: A new scene-consistent typography adversarial attack method (SceneTAP) is proposed, aimed at misleading large visual-language models (LVLMs) while maintaining visual naturalness.

SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow

Qingyuan Wang (Xidian University), Yinlin Hu (MagicLeap)

🎯 What it does: This paper proposes a plug-and-play 6D object pose refinement framework SCFlow2, which can iteratively improve the initial pose without the need for retraining or fine-tuning.

Schedule On the Fly: Diffusion Time Prediction for Faster and Better Image Generation

Zilyu Ye (Westlake University), Guo-Jun Qi (Westlake University)

🎯 What it does: This paper proposes a Time Prediction Diffusion Model (TPDM), which incorporates a Time Prediction Module (TPM) into the diffusion model to dynamically predict the noise level and total sampling steps for each instance, thereby improving image quality and sampling efficiency.

Science-T2I: Addressing Scientific Illusions in Image Synthesis

Jialuo Li (New York University), Saining Xie (New York University)

🎯 What it does: An expert-annotated adversarial dataset, Science-T2I, was constructed, and a SciScore reward model along with a two-stage (SFT+OFT) fine-tuning framework was proposed to enhance the authenticity and consistency of text-to-image models in scientific knowledge.

ScribbleLight: Single Image Indoor Relighting with Scribbles

Jun Myeong Choi (University of North Carolina), Roni Sengupta (University of North Carolina)

🎯 What it does: A single-image indoor scene relighting system based on doodles, ScribbleLight, has been designed and implemented, allowing users to achieve fine-grained lighting adjustments through simple bright/dark markings.

SCSA: A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer

Chunnan Shang (Zhejiang University), Xiangming Meng (Zhejiang University)

🎯 What it does: A plugin-based Semantic Continuous-Sparse Attention module (SCSA) is proposed to achieve arbitrary semantic style transfer.

SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures

Hui Liu (Tianjin University of Technology), Shengyong Chen (Tianjin University of Technology)

🎯 What it does: A lightweight structure-aware visual Mamba network, SCSegamba, is proposed for pixel-level segmentation of structural cracks.

SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models

Xiaofan Bai (Huazhong University of Science and Technology), Linchen Yu (Huazhong University of Science and Technology)

🎯 What it does: A hard label fingerprint detection method based on steep decision boundaries is proposed, which can detect tampering in cloud deep models in a black-box manner.

SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction

ZaiPeng Duan (Huazhong University of Science and Technology), Jie Ma (Huazhong University of Science and Technology)

🎯 What it does: A multi-modal 3D semantic occupancy prediction framework SDG-OCC is proposed, which combines camera and LiDAR information to achieve more accurate occupancy predictions.

Sea-ing in Low-light

Nisha Varghese (Indian Institute of Technology Madras), A. N. Rajagopalan (Indian Institute of Technology Madras)

🎯 What it does: Proposes SelfLUID-Net, a self-supervised learning framework that can simultaneously recover clear images and depth maps from a single low-light underwater image, achieving real-time inference.

SEAL: Semantic Attention Learning for Long Video Representation

Lan Wang (Michigan State University), Wen-Sheng Chu (Google)

🎯 What it does: The paper proposes the SEAL unified long video representation method, which achieves efficient and compressible representation in high-dimensional long videos by semantically decomposing them into three types of entities: scenes, objects, and actions, combined with attention learning.

SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation

Dekai Zhu (Technical University of Munich), Slobodan Ilic (Technical University of Munich)

🎯 What it does: A semantic part-aware latent point diffusion model called SeaLion is designed to generate high-quality 3D point clouds with point-level segmentation labels.

Search and Detect: Training-Free Long Tail Object Detection via Web-Image Retrieval

Mankeerat Sidhu (University of Illinois Urbana Champaign), Heng Ji (University of Illinois Urbana Champaign)

🎯 What it does: A training-free framework for long-tail open vocabulary object detection based on network retrieval of positive and negative samples (SearchDet) is proposed.

SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning

Ye Liu (Sun Yat-Sen University), Meng Yang (Sun Yat-Sen University)

🎯 What it does: Introducing the Semantic Complementary Prompt (SEC-Prompt) framework in few-shot class incremental learning.

SeCap: Self-Calibrating and Adaptive Prompts for Cross-view Person Re-Identification in Aerial-Ground Networks

Shining Wang (Northwestern Polytechnical University), Peng Wang (Northwestern Polytechnical University)

🎯 What it does: This paper proposes a Self-Calibrating Adaptive Prompt (SeCap) framework to address the challenge of cross-view person identification between drones and ground cameras.

Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis

Zexi Jia (WeChat AI Tencent Inc), Jie Zhou (WeChat AI Tencent Inc)

🎯 What it does: Proposes an AI-generated image detection method based on color distribution differences.

See Further When Clear: Curriculum Consistency Model

Yunpeng Liu (Institute of Computing Technology Chinese Academy of Sciences), Haihang You (Institute of Computing Technology Chinese Academy of Sciences)

🎯 What it does: A training framework based on a dynamic Curriculum Consistency Model (CCM) is proposed, which adaptively adjusts the teacher model's step size to maintain consistent knowledge differences under varying noise intensities, thereby achieving high-quality single-step sampling.

SeedVR: Seeding Infinity in Diffusion Transformer Towards Generic Video Restoration

Jianyi Wang (Nanyang Technological University), Lu Jiang (ByteDance)

🎯 What it does: We propose SeedVR, a general video restoration model based on diffusion transformers that can handle video inputs of arbitrary length and resolution.

SeeGround: See and Ground for Zero-Shot Open-Vocabulary 3D Visual Grounding

Rong Li (Hong Kong University of Science and Technology), Junwei Liang (Hong Kong University of Science and Technology)

🎯 What it does: The SeeGround framework is proposed to achieve zero-shot 3D visual localization, converting 3D scenes into rendered images and spatial text descriptions that can be processed by 2D visual language models (VLM), and aligning visual prompts with text.

Seeing A 3D World in A Grain of Sand

Yufan Zhang (George Mason University), Jinwei Ye (George Mason University)

🎯 What it does: A planar panoramic mirror system is designed to capture 360° snapshots of miniature scenes using eight pairs of flat mirrors on a nested pyramid.

Seeing Far and Clearly: Mitigating Hallucinations in MLLMs with Attention Causal Decoding

Feilong Tang (Monash University), Zongyuan Ge (Monash University)

🎯 What it does: A pluggable decoding strategy named FarSight is proposed, which suppresses the hallucination phenomenon in multimodal large language models (MLLMs) by adjusting the causal mask.

Seeing is Not Believing: Adversarial Natural Object Optimization for Hard-Label 3D Scene Attacks

Daizong Liu (Peking University), Wei Hu (Peking University)

🎯 What it does: A universal adversarial object optimization method is proposed for globally attacking 3D scene localization models using observable real object triggers under hard label conditions.

Seeing More with Less: Human-like Representations in Vision Models

Andrey Gizdov (Weizmann Institute of Science), Daniel Harari (Weizmann Institute of Science)

🎯 What it does: Introducing foveated sampling, which incorporates the distribution of the human eye's retina, into multimodal visual models to assess its impact on tasks such as VQA and object detection.

Seeing Speech and Sound: Distinguishing and Locating Audio Sources in Visual Scenes

Hyeonggon Ryu (Korea Advanced Institute of Science and Technology), Arda Senocak (Korea Advanced Institute of Science and Technology)

🎯 What it does: A unified model is proposed that can simultaneously locate speech and non-speech audio sources in a single image, supporting the learning and inference of mixed audio.

Seeing the Abstract: Translating the Abstract Language for Vision Language Models

Davide Talon (Fondazione Bruno Kessler), Yiming Wang

🎯 What it does: This study addresses the issue of the lack of expression of abstract language in visual language models within the fashion domain and proposes a training-free, model-agnostic abstract-to-concrete translation method (ACT) that significantly improves retrieval performance through language rewriting and representation offset compensation.

Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection

Gensheng Pei (Nanjing University of Science and Technology), Yazhou Yao (Nanjing University of Science and Technology)

🎯 What it does: This paper proposes CLIP-PGS, a progressive generation-selection masking strategy aimed at enhancing the training efficiency of CLIP in visual-language pre-training while preserving key information.

Seek Common Ground While Reserving Differences: Semi-Supervised Image-Text Sentiment Recognition

Wuyou Xia (Nankai University), Jufeng Yang (Nankai University)

🎯 What it does: A semi-supervised image-text sentiment recognition framework named SCRD is proposed, which significantly improves sentiment recognition performance in label-scarce environments by utilizing techniques such as feature decoupling (separating common and private features), unimodal classifiers, modality selection attention (MSeA), and pseudo-label filtering (PLF).

Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes

Aodi Li (University of Science and Technology of China), Shafei Wang (University of Science and Technology of China)

🎯 What it does: This paper proposes a Self-Feedback Training (SFT) framework that alternates between generating feedback signals and refining the loss landscape with soft labels to achieve cross-domain consistent flat minima, thereby enhancing domain generalization performance.

SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection

Haochen Li (Institute of Software), Ling Li (Institute of Software)

🎯 What it does: A domain-aware attention module SEEN-DA based on semantic entropy guidance is proposed for domain adaptive object detection.

SegAgent: Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories

Muzhi Zhu (Zhejiang University), Chunhua Shen (Zhejiang University)

🎯 What it does: This paper proposes the Human-like Annotation Trajectory task (HLMAT), which enables multi-modal LLM (MLLM) to achieve fine-grained pixel understanding through multi-step decision-making without introducing implicit pixel labels, and trains SegAgent based on this;

SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images

Kaiyu Li (Xi'an Jiaotong University), Zhi Wang (Xi'an Jiaotong University)

🎯 What it does: This paper proposes a training-free open-source vocabulary semantic segmentation method for remote sensing images—SegEarth-OV, specifically designed for remote sensing data, capable of performing multi-class pixel-level segmentation without the need for manual annotation.

SegMAN: Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation

Yunxiang Fu (University of Hong Kong), Yizhou Yu (University of Hong Kong)

🎯 What it does: This paper proposes a semantic segmentation network named SegMAN, which utilizes local attention and a two-dimensional state space model to achieve global context modeling and fine-grained feature encoding.

Segment Any Motion in Videos

Nan Huang (University of California Berkeley), Qianqian Wang (University of California Berkeley)

🎯 What it does: Combining long-term point trajectories with SAM2 to achieve unsupervised moving object segmentation, utilizing trajectory attention to encode trajectories, and then refining trajectory-level segmentation results into pixel-level masks through motion-semantic decoupling embedding.

Segment Any-Quality Images with Generative Latent Space Enhancement

Guangqian Guo (Northwestern Polytechnical University), Shan Gao (Northwestern Polytechnical University)

🎯 What it does: Introducing generative expansion in the latent space of the Segment Anything model, enhancing the segmentation robustness of low-quality images through single-step denoising.

Segment Anything, Even Occluded

Wei-En Tai (National Tsing Hua University), Hwann-Tzong Chen (National Taiwan University)

🎯 What it does: This study proposes a scalable invisible instance segmentation framework called SAMEO, which uses the Segment Anything model as a pluggable mask decoder and constructs a large-scale Amodal-LVIS dataset through synthetic occlusion data generation techniques.

Segment This Thing: Foveated Tokenization for Efficient Point-Prompted Segmentation

Tanner Schmidt (Meta Reality Labs), Richard Newcombe (Meta Reality Labs)

🎯 What it does: The Segment This Thing (STT) model is proposed to achieve efficient image segmentation with single-point prompts, utilizing foveated tokenization inspired by retinal perception.

Segmenting Maxillofacial Structures in CBCT Volumes

Federico Bolelli (University of Modena and Reggio Emilia), Costantino Grana (University of Modena and Reggio Emilia)

🎯 What it does: The largest publicly available CBCT dataset, ToothFairy2 (530 volume-level 3D annotations, containing 42 categories of facial structures), is proposed, and benchmark evaluations of various advanced 3D medical image segmentation models are conducted.

Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects

Weimin Qiu (University of California Merced), Meng Tang (University of California Merced)

🎯 What it does: A training-independent Self-Cross guidance method is proposed to address the issue of occlusion among multiple similar subjects by penalizing the overlap between aggregated self-attention and cross-attention during the inference process.

Self-Evolving Visual Concept Library using Vision-Language Critics

Atharva Sehgal (University of Texas), Swarat Chaudhuri (University of Texas)

🎯 What it does: This paper presents ESCHER, an iterative loop that automatically builds and improves a visual concept library through a VLM evaluator and an LLM generator, thereby enhancing the performance of concept bottleneck visual recognition systems.

Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning

Huiyi Wang (University of New South Wales), Dong Gong (University of New South Wales)

🎯 What it does: This paper proposes a self-expanding pre-training model adaptation method called SEMA for replay-free continual learning.

Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model

Jian Zhu (Nanjing University of Science and Technology), Zhihui Wei (Nanjing University of Science and Technology)

🎯 What it does: This paper proposes a self-learning adaptive residual-guided subspace diffusion model (ARGS-Diff) that achieves high-resolution hyperspectral image fusion using only low-resolution hyperspectral images and high-resolution multispectral images.

Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution

Shijun Shi (Jiangnan University), Kai Hu (Jiangnan University)

🎯 What it does: For real-world video super-resolution, a self-supervised ControlNet combined with a spatial-temporal Mamba has been proposed in the SCST framework, utilizing a pre-trained latent diffusion model to achieve high-quality, temporally consistent video enhancement.

Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency

Alan Baade (University of Texas at Austin), Changan Chen (Stanford University)

🎯 What it does: This paper studies a self-supervised cross-view consistency (Predictive Cycle Consistency, PCC) method to learn object correspondences in videos with extreme viewpoint changes and long time intervals.

Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration

Aocheng Li (Purdue University), Daniel Aliaga (Purdue University)

🎯 What it does: This paper proposes a self-supervised large-scale point cloud completion method that utilizes multi-center projection (MCOP) images to reconstruct missing structures in archaeological sites.

Self-Supervised Learning for Color Spike Camera Reconstruction

Yanchen Dong (Peking University), Tiejun Huang (Peking University)

🎯 What it does: A motion-guided reconstruction method for color spike cameras (CSC) is proposed, and a self-supervised network is built based on this to remove quantization noise, ultimately achieving high-quality color image recovery from Bayer-pattern spike flows.

Self-Supervised Spatial Correspondence Across Modalities

Ayush Shrivastava (University of Michigan), Andrew Owens (University of Michigan)

🎯 What it does: This paper proposes a self-supervised cross-modal pixel correspondence method that utilizes contrastive random walks and a global matching Transformer to achieve pixel-level matching for RGB, depth, thermal, photo-sketch, and cross-style images.

SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting

Gyeongjin Kang (Sungkyunkwan University), Eunbyung Park (Yonsei University)

🎯 What it does: This paper proposes SelfSplat, a general 3D Gaussian scattering model that is posture-free and does not rely on 3D priors, capable of directly reconstructing 3D scenes and synthesizing new views using only unlabelled multi-view images.

SemAlign3D: Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations

Krispin Wandel (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)

🎯 What it does: By combining monocular depth estimation with large model features, a 3D object class representation is constructed and aligned with RGB images during inference to achieve semantic correspondence.

Semantic and Expressive Variations in Image Captions Across Languages

Andre Ye (University of Washington), Ranjay Krishna (University of Washington)

🎯 What it does: This paper studies the differences in semantics and expression of image captions in different languages, and presents richer visual information through multilingual captions.

Semantic and Sequential Alignment for Referring Video Object Segmentation

Feiyu Pan (Shandong University), Xiankai Lu (Shandong University)

🎯 What it does: This paper proposes a framework called SSA based on semantic and sequence alignment for reference video object segmentation (RVOS);

Semantic Library Adaptation: LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation

Reza Qorbani (KTH), Matteo Poggi (University of Bologna)

🎯 What it does: This paper proposes the SemLA framework, which utilizes the LoRA adapter library and CLIP embeddings to achieve zero-training, real-time open vocabulary semantic segmentation domain adaptation.

Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition

Anqi Zhu (University of Melbourne), Qiuhong Ke (Monash University)

🎯 What it does: A semantic-guided cross-modal prompt learning framework (SCoPLe) is proposed, achieving zero-shot action recognition of skeletal sequences through dual-stream language prompts and skeletal prompts.

SemanticDraw: Towards Real-Time Interactive Content Creation from Image Diffusion Models

Jaerin Lee (Seoul National University), Kyoung Mu Lee (Seoul National University)

🎯 What it does: This paper presents SemanticDraw, a system that supports real-time interaction and allows for the generation of high-quality images through hand-drawn semantic masks controlling multiple text prompts on a canvas.

SemGeoMo: Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance

Peishan Cong (ShanghaiTech University), Xiangyu Yue (Chinese University of Hong Kong)

🎯 What it does: This paper proposes the SemGeoMo method, which can generate complete human interaction actions and corresponding text descriptions with semantic coherence and geometric accuracy under the condition of given dynamic interactive target point cloud sequences.

Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining

Shangquan Sun (Institute of Information Engineering Chinese Academy of Sciences), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)

🎯 What it does: A dual-branch spatiotemporal state space model VDMamba is proposed, which implements video de-raining using dynamic stacked filters and semi-supervised stacked loss.

SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation

Hritam Basak (Stony Brook University), Zhaozheng Yin (Stony Brook University)

🎯 What it does: This paper proposes the SemiDAViL framework to study the problem of semi-supervised domain adaptation (SSDA) for semantic segmentation.

SemiETS: Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting

Dongliang Luo (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)

🎯 What it does: A semi-supervised end-to-end text recognition framework called SemiETS is proposed, which can utilize unlabeled images to enhance text recognition performance.

Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation

Tianran Chen (Harbin Institute of Technology), Yunming Ye (Harbin Institute of Technology)

🎯 What it does: A dynamic rank adaptation method based on the distribution of sensitive parameter rows and columns is proposed for parameter-efficient fine-tuning.

Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering

Zhen Yang (University of Adelaide), Qingming Huang (Institute of Computing Technology, Chinese Academy of Sciences)

🎯 What it does: A question-answering aware image descriptor QACap is proposed, which guides questions to image feature extraction and caption generation, thereby generating descriptions related to the questions as context for LLM.

Seq2Time: Sequential Knowledge Transfer for Video LLM Temporal Grounding

Andong Deng (University of Central Florida), Ziyan Wu (United Imaging Intelligence)

🎯 What it does: Proposes the Seq2Time framework, which self-supervises the generation of time labels from image sequences and short video clips to enhance the temporal awareness of video LLMs.

SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model

Chunlin Yu (ShanghaiTech University), Jingya Wang (ShanghaiTech University)

SeqMvRL: A Sequential Fusion Framework for Multi-view Representation Learning

Ren Wang (Shandong University), Wenjia Meng (Shandong University)

🎯 What it does: This paper proposes SeqMvRL, a sequential fusion framework based on reinforcement learning for multi-view representation learning, which can adaptively select view sequences and gradually fuse them to generate high-quality features.

SerialGen: Personalized Image Generation by First Standardization Then Personalization

Cong Xie (Baidu Inc), Zhenpeng Zhan

🎯 What it does: A two-stage personalized image generation framework called SerialGen is proposed, which first standardizes the reference image and then generates personalized images based on the standardized image.

SeriesBench: A Benchmark for Narrative-Driven Drama Series Understanding

Chenkai Zhang (Beihang University), Yunhong Wang (Beihang University)

🎯 What it does: Proposes the SeriesBench benchmark and the PC-DCoT framework for evaluating and enhancing multimodal large language models' understanding of long-form narratives in episodic dramas.

SET: Spectral Enhancement for Tiny Object Detection

Huixin Sun (Beihang University), Baochang Zhang (Beihang University)

🎯 What it does: The Spectral Enhancement for Tiny objects (SET) method is proposed for sparse object detection tasks, combining background smoothing and adversarial perturbation injection to enhance the frequency domain features and detection performance of small targets.

Seurat: From Moving Points to Depth

Seokju Cho (KAIST), Joon-Young Lee (Adobe Research)

🎯 What it does: Predict relative depth changes from the 2D point trajectories of monocular video, generating smooth and accurate 3D depth sequences.

SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding

Yangliu Hu (Huazhong University of Science and Technology), Wei Yang (Zhejiang University)

🎯 What it does: This paper proposes a self-supervised fragment tuning (SFT) method to enhance the performance of video large language models in fine-grained spatiotemporal understanding, and constructs a multi-level evaluation benchmark called FineVidBench.

SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement

Mark Boss (Stability AI), Varun Jampani (Stability AI)

🎯 What it does: This paper proposes a system named SF3D, which can quickly generate high-quality 3D meshes with UV unwrapping, material properties, and de-lighting from a single image in 0.3 seconds.

SFDM: Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images

Daisheng Jin (Nanyang Technological University), Ying He (Nanyang Technological University)

🎯 What it does: A two-stage sparse view 3D face reconstruction and decomposition method is proposed - SFDM;

SfM-Free 3D Gaussian Splatting via Hierarchical Training

Bo Ji (National University of Singapore), Angela Yao (National University of Singapore)

🎯 What it does: A 3D Gaussian spraying method without SfM preprocessing is proposed, which combines basic 3DGS models of different scene segments into a complete scene model using hierarchical training.

SGC-Net: Stratified Granular Comparison Network for Open-Vocabulary HOI Detection

Xin Lin, Zhili Zhou

🎯 What it does: This paper proposes an image classification network that integrates attention mechanisms with multi-scale feature fusion;

SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction

Xinran Yang (Nanjing University), Junyuan Xie (Nanjing University)

🎯 What it does: Generate Spherical Gaussians from 2D image edge information and use them to extract 3D parametric curves, achieving efficient 3D curve reconstruction.

SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion

Xiyue Guo (Zhejiang University), Guofeng Zhang (Zhejiang University)

🎯 What it does: This paper proposes SGFormer, a 3D semantic scene completion framework that integrates satellite images with ground camera views, addressing the occlusion bottleneck that arises from relying solely on ground perspectives.

SGSST: Scaling Gaussian Splatting Style Transfer

Bruno Galerne (University of Orléans), Jean-Michel Morel (University of Orléans)

🎯 What it does: This study investigates the method of transferring 2D style images to 3D high-resolution Gaussian splatting (3DGS) scenes.

Shading Meets Motion: Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion

Guoyu Lu (Binghamton University)

🎯 What it does: A self-supervised framework called Shading-SfM-Net is proposed to simultaneously learn shape from shading (SfS) and structure from motion (SfM) for low-texture indoor 3D reconstruction.

Shadow Generation Using Diffusion Model with Geometry Prior

Haonan Zhao (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)

🎯 What it does: The research uses geometric priors to guide the diffusion model in generating synthetic image shadows, significantly improving the geometric quality of the shadows.

Shape Abstraction via Marching Differentiable Support Functions

Sunkyung Park (Seoul National University), Dongjun Lee (Seoul National University)

🎯 What it does: A method for shape abstraction using Differentiable Support Functions (DSF) is proposed, which abstracts complex geometries into a small number of differentiable convex bodies.

Shape and Texture: What Influences Reliable Optical Flow Estimation?

Libo Long (University of Ottawa), Jochen Lang (University of Ottawa)

🎯 What it does: This paper constructs the Flow-R dataset by modifying the shape and texture of target objects and adding unseen objects to the original KITTI images, aimed at evaluating the robustness of optical flow estimation.

Shape My Moves: Text-Driven Shape-Aware Synthesis of Human Motions

Ting-Hsuan Liao (University of Maryland), Uttaran Bhattacharya (Adobe Research)

🎯 What it does: Simultaneously generate corresponding human body shape parameters and shape-aware 3D action sequences from natural language descriptions.

ShapeShifter: 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion

Nissim Maruani (Inria Universit Côte d'Azur), Mathieu Desbrun (Inria)

🎯 What it does: A 3D single-instance generation model named ShapeShifter is proposed, which can quickly generate high-quality and diverse deformations from a single reference model and supports interactive editing.

ShapeWords: Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts

Dmitry Petrov (University of Massachusetts Amherst), Evangelos Kalogerakis (University of Massachusetts Amherst)

🎯 What it does: The ShapeWords method is proposed, which embeds 3D shape information into CLIP word vectors and combines it with text prompts to achieve diffusion-based text-to-image generation guided by 3D shapes.

Sharp-It: A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation

Yiftach Edelstein, Lihi Zelnik-Manor

🎯 What it does: This paper presents Sharp-It, a multi-view to multi-view diffusion model that enhances the details and textures of multi-view images rendered by low-quality 3D generative models (such as Shap-E), thereby generating high-quality editable 3D assets.

SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation

Duc-Hai Pham (Movian AI), Rang Nguyen (Movian AI)

🎯 What it does: We propose SharpDepth, a zero-shot metric depth sharpening method based on diffusion models, which can maintain both absolute scale accuracy and boundary clarity without the need for real depth labels.