CVPR 2024 Papers — Page 22
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2716 papers
SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing
Zeyinzi Jiang (Alibaba Group), Jingfeng Zhang (Alibaba Group)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: The SCEdit method is proposed, which achieves efficient fine-tuning and controllable image generation of diffusion models by incorporating a lightweight SC-Tuner object into the skip connections of U-Net.
Scene Adaptive Sparse Transformer for Event-based Object Detection
Yansong Peng (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
Object DetectionTransformerImage
🎯 What it does: A sparse Transformer architecture for event cameras (SAST) is proposed, which significantly reduces computational load while maintaining detection performance through window-tag co-sparsification.
Scene-adaptive and Region-aware Multi-modal Prompt for Open Vocabulary Object Detection
Xiaowei Zhao (Beihang University), Zhide Liu (Beihang University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the Scene-adaptive and Region-aware Multi-modal Prompt (SAMP) framework for open vocabulary object detection, which efficiently transfers the knowledge of pre-trained vision-language models to detection networks.
SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes
Alexandros Delitzas (ETH Zurich), Francis Engelmann (Google)
SegmentationPose EstimationConvolutional Neural NetworkTextPoint CloudBenchmark
🎯 What it does: This paper presents the SceneFun3D dataset, which provides fine annotations, motion parameters, and natural language task descriptions for functional interactive elements in 14.8k high-resolution indoor scenes. Based on this, three new benchmark tasks are designed: functional segmentation, task-driven empowered localization, and 3D action estimation.
SceneTex: High-Quality Texture Synthesis for Indoor Scenes via Diffusion Priors
Dave Zhenyu Chen (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: This paper presents SceneTex, which utilizes a diffusion model prior from depth to image to achieve high-quality, style-consistent texture synthesis on 3D indoor scene meshes using multi-resolution implicit texture fields and cross-attention decoders.
SchurVINS: Schur Complement-Based Lightweight Visual Inertial Navigation System
Yunfei Fan (ByteDance), Guidong Wang (ByteDance)
OptimizationComputational EfficiencySimultaneous Localization and MappingVideo
🎯 What it does: A lightweight visual-inertial navigation system based on Schur complement (SchurVINS) is proposed, which utilizes EKF for joint estimation of attitude and landmarks within a sliding window, significantly reducing computational complexity.
SCINeRF: Neural Radiance Fields from a Snapshot Compressive Image
Yunhao Li (Zhejiang University), Peidong Liu (Westlake University)
RestorationGenerationData SynthesisCompressionNeural Radiance FieldImage
🎯 What it does: Recover 3D scene representation from a single compressed image and generate multi-view consistent high frame rate images based on that representation.
SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation
Zhixuan Liu, Jean Oh
GenerationDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: A self-contrast fine-tuning method called SCoFT is proposed, based on a small cross-cultural dataset, to improve the cultural expression accuracy and diversity of the stable diffusion model in image generation.
Score-Guided Diffusion for 3D Human Recovery
Anastasis Stathopoulos (Rutgers University), Dimitris Metaxas (Rutgers University)
Pose EstimationDiffusion modelScore-based ModelImageVideo
🎯 What it does: This paper proposes ScoreHMR, an iterative refinement method based on diffusion models for the inverse problem of 3D human pose and shape reconstruction.
ScoreHypo: Probabilistic Human Mesh Estimation with Hypothesis Scoring
Yuan Xu (Peking University), Yizhou Wang (Peking University)
GenerationPose EstimationTransformerDiffusion modelImageMesh
🎯 What it does: This work proposes the ScoreHypo framework, which combines the multi-hypothesis 3D human mesh generator HypoNet with a general hypothesis scorer ScoreNet, enabling the generation of multiple feasible estimates from a single image and selecting high-quality results.
SCULPT: Shape-Conditioned Unpaired Learning of Pose-dependent Clothed and Textured Human Meshes
Soubhik Sanyal (Max Planck Institute for Intelligent Systems), Timo Bolkart (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisPose EstimationGenerative Adversarial NetworkImageMesh
🎯 What it does: A generative model named SCULPT has been developed, which can synthesize textured 3D human meshes based on attributes such as pose, clothing type, and color; this model supports free pose deformation and can directly output explicit meshes suitable for games and rendering engines.
Sculpt3D: Multi-View Consistent Text-to-3D Generation with Sparse 3D Prior
Cheng Chen (Astar), Fayao Liu (Astar)
GenerationData SynthesisRetrievalDiffusion modelNeural Radiance FieldTextPoint CloudMesh
🎯 What it does: This paper proposes a method for text-driven 3D generation by retrieving template objects from external 3D datasets and combining them with 2D diffusion models, aiming to enhance the geometric consistency and multi-view appearance consistency of the generated objects.
Sculpting Holistic 3D Representation in Contrastive Language-Image-3D Pre-training
Yipeng Gao (University of California Santa Cruz), Yuyin Zhou (Sun Yat-sen University)
RecognitionRepresentation LearningContrastive LearningImageTextPoint Cloud
🎯 What it does: This paper proposes MixCon3D, which constructs a unified 3D object-level representation by fusing multi-view RGB and point cloud features, and achieves open-source 3D understanding through tri-modal (image, text, point cloud) contrastive learning.
SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer
Rui Zhu (Chinese University of Hong Kong), Chang Wen Chen (Hong Kong Polytechnic University)
GenerationKnowledge DistillationTransformerDiffusion modelImage
🎯 What it does: A novel Diffusion Transformer (SD-DiT) is designed, which efficiently trains generative models through a teacher-student framework of self-supervised discriminative knowledge distillation for masked modeling.
SD2Event:Self-supervised Learning of Dynamic Detectors and Contextual Descriptors for Event Cameras
Yuan Gao (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Object DetectionRepresentation LearningAgentic AIContrastive LearningImageVideo
🎯 What it does: This paper proposes a self-supervised learning-based event camera keypoint detection framework called SD2Event, which can simultaneously learn dynamic keypoint detectors and context descriptors, suitable for unlabeled event data.
SD4Match: Learning to Prompt Stable Diffusion Model for Semantic Matching
Xinghui Li (University of Oxford), Victor Adrian Prisacariu (University of Oxford)
RecognitionRetrievalPrompt EngineeringDiffusion modelImage
🎯 What it does: Proposes SD4Match, which optimizes the prompts of Stable Diffusion to generate more distinctive features in semantic matching tasks;
SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection
Junsu Kim (Ulsan National Institute of Science and Technology), Seungryul Baek (Ulsan National Institute of Science and Technology)
Object DetectionGenerationData SynthesisKnowledge DistillationDiffusion modelImage
🎯 What it does: A method called Stable Diffusion-based Deep Generative Replay (SDDGR) is proposed to address the problem of catastrophic forgetting in class-incremental object detection.
SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation
Sichen Chen (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
Pose EstimationKnowledge DistillationTransformerImage
🎯 What it does: A self-distillation framework based on multi-cycle forward, SDPose, is proposed to specifically enhance the performance of small Transformers in human pose estimation tasks.
SDSTrack: Self-Distillation Symmetric Adapter Learning for Multi-Modal Visual Object Tracking
Xiaojun Hou (Zhejiang University), Yong Liu (Zhejiang University)
Object TrackingKnowledge DistillationTransformerMultimodality
🎯 What it does: A multi-modal visual object tracking framework called SDSTrack is proposed, which efficiently utilizes a pre-trained RGB tracker and integrates information from depth, thermal, and event modalities through self-distillation symmetric adaptation.
SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects
Abhinav Kumar, Xiaoming Liu
Object DetectionSegmentationAutonomous DrivingImage
🎯 What it does: This study investigates the generalization problem of monocular 3D detection on large objects and proposes SeaBird, which enhances detection accuracy through BEV segmentation and Dice loss.
Seamless Human Motion Composition with Blended Positional Encodings
German Barquero (Universitat de Barcelona), Cristina Palmero (Universitat de Barcelona)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideoText
🎯 What it does: A FlowMDM based on diffusion models is proposed for generating seamless and continuous long-term human motion in a single instance, which can be controlled according to multiple text descriptions.
SEAS: ShapE-Aligned Supervision for Person Re-Identification
Haidong Zhu (University of Southern California), Ram Nevatia (University of Southern California)
RecognitionRetrievalConvolutional Neural NetworkImageVideo
🎯 What it does: Utilizing 3-D human shapes as pixel-level supervision to enhance the appearance feature encoder during the training phase, thereby improving pedestrian recognition performance for both single frames and videos.
SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation
Yamei Chen (Technical University of Munich), Benjamin Busam (Technical University of Munich)
Pose EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes SecondPose, a method for category-level 9D object pose estimation achieved through dual-stream feature fusion.
SED: A Simple Encoder-Decoder for Open-Vocabulary Semantic Segmentation
Bin Xie (Tianjin University), Yanwei Pang (Tianjin University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A concise encoder-decoder framework SED is proposed for open vocabulary semantic segmentation.
SeD: Semantic-Aware Discriminator for Image Super-Resolution
Bingchen Li (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
RestorationSuper ResolutionGenerative Adversarial NetworkImage
🎯 What it does: A semantic-aware discriminator (SeD) is proposed to improve GAN-based image super-resolution (SR) models, enabling them to generate finer-grained textures that are semantically consistent.
See Say and Segment: Teaching LMMs to Overcome False Premises
Tsung-Han Wu (University of California, Berkeley), Trevor Darrell (University of California, Berkeley)
Object DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A multimodal model called SESAME is proposed, which simultaneously possesses detection (see), response (say), and segmentation (segment) capabilities, addressing the issue of false premise in reasoning segmentation regarding non-existent objects.
SEED-Bench: Benchmarking Multimodal Large Language Models
Bohao Li (Tencent AI Lab), Ying Shan (Tencent AI Lab)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A multi-modal large language model (MLLM) evaluation benchmark named SEED-Bench has been constructed, covering multi-level capabilities from text understanding to image generation, and evaluated across 27 dimensions using 24,000 multiple-choice questions.
Seeing and Hearing: Open-domain Visual-Audio Generation with Diffusion Latent Aligners
Yazhou Xing (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisPrompt EngineeringDiffusion modelImageVideoMultimodalityAudio
🎯 What it does: This paper proposes an optimized cross-modal generation framework based on a pre-trained unimodal diffusion model and the ImageBind multimodal embedding space, achieving joint and mutual generation of open-domain visual and audio content.
Seeing Motion at Nighttime with an Event Camera
Haoyue Liu (Huazhong University of Science and Technology), Luxin Yan (Huazhong University of Science and Technology)
RestorationObject DetectionOptical FlowImageVideoBenchmark
🎯 What it does: This study proposes a nighttime dynamic scene imaging method based on event cameras and constructs the first real low-light event and high-quality image alignment dataset, RLED.
Seeing the Unseen: Visual Common Sense for Semantic Placement
Ram Ramrakhya (Georgia Institute of Technology), Luca Weihs (Allen Institute of AI)
Object DetectionSegmentationRobotic IntelligenceConvolutional Neural NetworkSupervised Fine-TuningDiffusion modelSimultaneous Localization and MappingImage
🎯 What it does: Proposes the Semantic Placement (SP) task: Given an image and the name of a target object, predict the semantic mask where the object can be placed; constructs a large-scale unlabeled dataset through an automated data generation pipeline (image retrieval → object detection → segmentation → object removal (inpainting) → filtering and enhancement) to train the CLIP-UNet model; and applies this model in applications such as assistant robots and AR rendering.
Seeing the World through Your Eyes
Hadi Alzayer (University of Maryland), Jia-Bin Huang (University of Maryland)
RestorationPose EstimationOptimizationNeural Radiance FieldImage
🎯 What it does: Recover the radiance field of the scene seen by the observer from a sequence of faces captured by a fixed camera using the reflection of the human eye.
Seeing Unseen: Discover Novel Biomedical Concepts via Geometry-Constrained Probabilistic Modeling
Jianan Fan (University of Sydney), Weidong Cai (University of Sydney)
ClassificationSegmentationAnomaly DetectionDrug DiscoveryTransformerContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A geometric constraint-based probabilistic modeling method is proposed for non-i.i.d. and severely imbalanced medical imaging data, enabling the automatic discovery of unknown medical concepts (cells, lesions, diseases, etc.).
SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution
Rongyuan Wu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: This paper proposes a method called SeeSR that utilizes semantic prompts (hard labels and soft features) to control pre-trained text-image diffusion models for real-world image super-resolution.
Seg2Reg: Differentiable 2D Segmentation to 1D Regression Rendering for 360 Room Layout Reconstruction
Cheng Sun (NVIDIA), Hwann-Tzong Chen (Aeolus Robotics)
SegmentationDepth EstimationConvolutional Neural NetworkNeural Radiance FieldImage
🎯 What it does: This paper proposes Seg2Reg, which transforms the 2D density field of 360° panoramic images into 1D depth regression through differentiable volume rendering, achieving end-to-end reconstruction of indoor room layouts.
Segment and Caption Anything
Xiaoke Huang (Shenzhen International Graduate School Tsinghua University), Zicheng Liu (Advanced Micro Devices)
SegmentationGenerationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: By adding a lightweight query-based feature mixer to SAM, the regional features are aligned to the pre-trained language model, enabling the automatic generation of captions for image regions.
Segment Any Event Streams via Weighted Adaptation of Pivotal Tokens
Zhiwen Chen, Jinjian Wu
RetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a new cross-modal retrieval framework for image retrieval tasks, achieving more efficient retrieval matching through joint encoding of image features and text descriptions.
Segment Every Out-of-Distribution Object
Wenjie Zhao (Harvard University), Yunhui Guo (University of Texas at Dallas)
Object DetectionSegmentationAnomaly DetectionPrompt EngineeringImage
🎯 What it does: Maps anomaly scores to box prompts, using a promptable segmentation model (such as SAM) to directly generate high-quality OoD object masks, achieving OoD detection without the need for thresholds.
Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model
Dian Zheng (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes a selective 'hourglass' mapping strategy based on diffusion models, utilizing shared distribution mapping and strong conditional guidance to achieve unified image restoration.
Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition
Filip Ilic (Graz University of Technology), Richard P. Wildes (Graz University of Technology)
RecognitionSafty and PrivacyExplainability and InterpretabilityTransformerOptical FlowVideo
🎯 What it does: A selective, interpretable, and temporally consistent privacy attribute masking method based on handcrafted templates is proposed to protect personal privacy while maintaining action recognition performance.
Selective Nonlinearities Removal from Digital Signals
Krzysztof A. Maliszewski (University of Canterbury), Sylwia M. Kolenderska (Nicolaus Copernicus University)
RestorationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImageComputed Tomography
🎯 What it does: This paper proposes a technique for removing different orders of nonlinearity in digital signals such as Optical Coherence Tomography (OCT) using neural networks, capable of separately removing second or third-order nonlinearity.
Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching
Xianqi Wang (Huazhong University of Science and Technology), Xin Yang (Huazhong University of Science and Technology)
Depth EstimationOptimizationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: A Selective-Stereo framework is proposed, which incorporates Selective Recurrent Unit (SRU) and Contextual Spatial Attention (CSA) modules into traditional iterative stereo matching networks to adaptively fuse information of different frequencies (high-frequency details and low-frequency smoothness), thereby improving the quality of dense disparity estimation.
Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization
Jimyeong Kim (Seoul National University), Wonjong Rhee (Seoul National University)
GenerationData SynthesisOptimizationLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A text description strategy named Selectively Informative Description (SID) is proposed to reduce the embedding entanglement problem caused by non-target objects in reference images during personalized text-to-image generation.
Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution
Qingping Zheng (Northwestern Polytechnical University), Hang Xu (Huawei Noah's Ark Lab)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A training-free Self-Adaptive Reality-Guided Diffusion (SARGD) framework is proposed for real-time detection and elimination of artifacts during the diffusion-based super-resolution process, while enhancing image details and realism.
Self-Calibrating Vicinal Risk Minimisation for Model Calibration
Jiawei Liu (Australian National University), Nick Barnes (Australian National University)
ClassificationSegmentationConvolutional Neural NetworkImage
🎯 What it does: A self-calibrating Vicinal Risk Minimization (SCVRM) method is proposed, which enhances model calibration of deep networks in dense binary classification tasks by generating augmented images with distance-decaying labels near the training samples.
Self-correcting LLM-controlled Diffusion Models
Tsung-Han Wu, Trevor Darrell (University of California Berkeley)
Object DetectionGenerationTransformerLarge Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: A self-correcting text-to-image generation framework has been designed and implemented, which iteratively corrects generated images through LLM-driven detection and control.
Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation
Hang Li (LMU Munich), Jindong Gu (University of Oxford)
GenerationDiffusion modelImage
🎯 What it does: A self-supervised method is proposed to discover and learn interpretable latent directions in the bottleneck layer h-space of diffusion models, used to represent arbitrary concepts (such as gender, safety, etc.), and to achieve fair generation, safe generation, and enhanced accountability in text-to-image generation by adding corresponding vectors during inference.
Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors
Nicolae-C?t?lin Ristea, Mubarak Shah
Anomaly DetectionKnowledge DistillationTransformerAuto EncoderVideo
🎯 What it does: A lightweight self-distillation masked autoencoder is proposed for video anomaly detection, combining motion gradient weighting, teacher-student decoding, and synthetic anomaly enhancement.
Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations
Kewei Wang (Huazhong University of Science and Technology), Guosheng Lin (Nanyang Technological University)
Autonomous DrivingContrastive LearningPoint Cloud
🎯 What it does: Utilizing unlabeled LiDAR sequences, pseudo-motion labels are generated through optimal transport matching, and based on this, a class-agnostic motion prediction model is trained.
Self-supervised Debiasing Using Low Rank Regularization
Geon Yeong Park (Korea Advanced Institute of Science and Technology), Sang Wan Lee (Korea Advanced Institute of Science and Technology)
ClassificationRepresentation LearningContrastive LearningImageElectronic Health Records
🎯 What it does: This paper proposes a self-supervised debiasing framework called DeFund, which utilizes low-rank regularization to train a bias encoder that identifies and weights bias-conflicted samples, thereby enhancing the model's generalization performance under unsupervised or low-label conditions.
Self-Supervised Dual Contouring
Ramana Sundararaman (Ecole Polytechnique), Maks Ovsjanikov (Ecole Polytechnique)
GenerationData SynthesisConvolutional Neural NetworkMesh
🎯 What it does: This paper proposes a self-supervised Dual Contouring method (SDC) that directly predicts mesh vertices from SDF grids through two types of geometric consistency losses, eliminating the reliance on QEF solving and manually trained data. This self-supervised framework is applied for the regularization of Deep Implicit Networks (DIN) and for end-to-end reconstruction from single-view images to meshes.
Self-Supervised Facial Representation Learning with Facial Region Awareness
Zheng Gao (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
RecognitionRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A facial representation learning framework F R A is proposed, which utilizes self-supervised methods to learn globally and locally (facial region) consistent representations, with ResNet as the general backbone network.
Self-Supervised Multi-Object Tracking with Path Consistency
Zijia Lu (Northeastern University), Davide Modolo (Amazon Web Services)
Object TrackingTransformerContrastive LearningVideo
🎯 What it does: The concept of Path Consistency is proposed, and a model capable of robust multi-object tracking over long time intervals is trained using its self-supervised loss.
Self-Supervised Representation Learning from Arbitrary Scenarios
Zhaowen Li (Institute of Automation, Chinese Academy of Science), Jinqiao Wang (Institute of Automation, Chinese Academy of Science)
ClassificationObject DetectionSegmentationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: A self-supervised learning framework for arbitrary scenes, ASL, is proposed, utilizing MAE encoders and decoders, and generating dual branches through feature enhancement to replace traditional global contrastive learning with patch-level contrastive learning, resolving the conflict between MAE and CL.
Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement
Zaid Khan (Northeastern University), Manmohan Chandraker (NEC Laboratories America)
Object DetectionTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageTextMultimodality
🎯 What it does: This work proposes VisReP, a framework based on reinforcement self-training that utilizes coarse-grained rewards obtained from visual program execution to train large language models (LLMs) to generate more accurate visual programs.
SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction
Yuanhui Huang (Tsinghua University), Jiwen Lu (Tsinghua University)
SegmentationDepth EstimationAutonomous DrivingNeural Radiance FieldVideo
🎯 What it does: This paper proposes SelfOcc, which utilizes a self-supervised approach based solely on video sequences to infer 3D occupancy predictions in BEV/TPV space from multi-camera RGB images, thus avoiding the need for 3D annotations.
SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation
Vinkle Srivastav (University of Strasbourg), Nicolas Padoy (University of Strasbourg)
Pose EstimationImage
🎯 What it does: A fully self-supervised multi-camera multi-person 3D pose estimation method called SelfPose3d is proposed, which utilizes 2D pseudo-poses and multi-view geometric constraints to achieve 3D pose reconstruction without the need for 2D or 3D real annotations.
Semantic Human Mesh Reconstruction with Textures
Xiaoyu Zhan (Nanjing University), Wenping Wang (Texas A&M University)
RestorationSegmentationGenerationDiffusion modelImageMesh
🎯 What it does: This paper presents SHERT, which utilizes detailed 3D surfaces or single images along with the SMPL-X prior to generate complete high-quality human meshes with semantic labels, skin weights, and animation capabilities, and enables text-driven texture repair and generation based on this.
Semantic Line Combination Detector
Jinwon Ko, Chang-Su Kim
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new computer vision method aimed at improving the accuracy of image recognition.
Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment
Alvi Md Ishmam (Virginia Tech), Christopher Thomas (Virginia Tech)
Representation LearningAdversarial AttackTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A Semantic Shield defense scheme is proposed for contrastive learning visual-language models to prevent backdoor and poisoning attacks.
Semantic-Aware Multi-Label Adversarial Attacks
Hassan Mahmood (Northeastern University), Ehsan Elhamifar (Northeastern University)
Adversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a semantic-aware multi-label adversarial attack framework (GMLA) that can generate effective attacks while maintaining label semantic consistency.
Semantic-aware SAM for Point-Prompted Instance Segmentation
Zhaoyang Wei (University of Chinese Academy of Sciences), Zhenjun Han (University of Chinese Academy of Sciences)
Object DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: Proposes SAPNet, an end-to-end instance segmentation framework based on single-point prompts.
Semantically-Shifted Incremental Adapter-Tuning is A Continual ViTransformer
Yuwen Tan (Huazhong University of Science and Technology), Yongbin Li (Alibaba Group)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningImage
🎯 What it does: Research on class-incremental learning based on pre-trained Vision Transformers is conducted, proposing a parameter-free shared Adapter incremental fine-tuning combined with semantic drift estimation to retrain a unified classifier.
Semantics Distortion and Style Matter: Towards Source-free UDA for Panoramic Segmentation
Xu Zheng (Hong Kong University of Science and Technology), Lin Wang (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationTransformerImage
🎯 What it does: This paper proposes a Source-Free Unsupervised Domain Adaptation (SFUDA) method for transferring a pre-trained pinhole camera segmentation model to the semantic segmentation task of 360° panoramic images.
Semantics-aware Motion Retargeting with Vision-Language Models
Haodong Zhang (Zhejiang University), Rong Xiong (Zhejiang University)
Graph Neural NetworkTransformerVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes a semantic-aware action redirection method guided by a visual-language model, rendering 3D actions into differentiable images and obtaining action semantics through guided visual question answering.
SemCity: Semantic Scene Generation with Triplane Diffusion
Jumin Lee (Korea Advanced Institute of Science and Technology), Sung-Eui Yoon (Korea Advanced Institute of Science and Technology)
GenerationOptimizationDiffusion modelPoint Cloud
🎯 What it does: This paper proposes SemCity, which generates realistic outdoor semantic scenes based on a three-plane diffusion model, and implements scene filling, extension, and semantic scene completion optimization through three-plane operations.
SeMoLi: What Moves Together Belongs Together
Jenny Seidenschwarz (Technical University of Munich), Laura Leal-Taixe
Object DetectionSegmentationAutonomous DrivingGraph Neural NetworkSimultaneous Localization and MappingOptical FlowPoint Cloud
🎯 What it does: Developed an unsupervised point cloud instance segmentation and pseudo-label generation framework based on motion patterns, named SeMoLi.
SeNM-VAE: Semi-Supervised Noise Modeling with Hierarchical Variational Autoencoder
Dihan Zheng (Tsinghua University), Chenglong Bao (Tsinghua University)
RestorationData SynthesisSuper ResolutionAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A semi-supervised noise modeling framework SeNM-VAE is proposed, which generates high-quality denoising/super-resolution training samples using a small amount of paired data and a large amount of unpaired data.
Separate and Conquer: Decoupling Co-occurrence via Decomposition and Representation for Weakly Supervised Semantic Segmentation
Zhiwei Yang (Fudan University), Zhijian Song (Fudan University)
SegmentationKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: A single-stage weakly supervised semantic segmentation framework called SeCo is proposed, which separates co-occurring objects through image patching and category label partitioning, and enhances feature representation using multi-granularity contrastive learning and knowledge distillation to address the misactivation problem caused by co-occurrence.
Separating the "Chirp" from the "Chat": Self-supervised Visual Grounding of Sound and Language
Mark Hamilton (Massachusetts Institute of Technology), William T. Freeman (Massachusetts Institute of Technology)
SegmentationRetrievalTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: Using unsupervised contrastive learning and relying solely on video training, a high-resolution, semantically aligned dual encoder was constructed to achieve local alignment and semantic localization of audio and vision.
Sequential Modeling Enables Scalable Learning for Large Vision Models
Yutong Bai (Johns Hopkins University), Alexei A. Efros (University of California Berkeley)
SegmentationDepth EstimationTransformerVision Language ModelImageVideo
🎯 What it does: This study proposes a unified format of 'visual sentences' and trains an autoregressive large visual model (LVM) that can predict the next visual token using large-scale unsupervised data.
SfmCAD: Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations
Pu Li (Institute of Automation Chinese Academy of Sciences), Dong-Ming Yan (Institute of Automation Chinese Academy of Sciences)
GenerationData SynthesisConvolutional Neural NetworkPoint CloudMesh
🎯 What it does: This paper proposes an unsupervised SfmCAD network that automatically decomposes voxel input into editable CAD operation sequences (Sketch + Path) and reconstructs 3D shapes.
SFOD: Spiking Fusion Object Detector
Yimeng Fan (Tianjin University), Wenrui Lu (Tianjin University)
Object DetectionSpiking Neural NetworkImage
🎯 What it does: For target detection with event cameras, we propose the Spiking Fusion Object Detector (SFOD), which implements multi-scale feature fusion and completes the target detection task within the SNN framework.
SG-BEV: Satellite-Guided BEV Fusion for Cross-View Semantic Segmentation
Junyan Ye (Sun Yat-Sen University), Weijia Li (Sun Yat-Sen University)
SegmentationTransformerImage
🎯 What it does: By combining satellite images with street view images, fine-grained semantic segmentation of architectural attributes across different perspectives is achieved.
SG-PGM: Partial Graph Matching Network with Semantic Geometric Fusion for 3D Scene Graph Alignment and Its Downstream Tasks
Yaxu Xie (German Research Center for Artificial Intelligence), Didier Stricker (German Research Center for Artificial Intelligence)
Object DetectionSegmentationGraph Neural NetworkPoint CloudGraph
🎯 What it does: This study addresses the problem of 3D scene graph alignment and proposes a partial graph matching method based on graph neural networks called SG-PGM. It combines semantic and geometric fusion to achieve node matching, which is then used for downstream tasks such as overlap detection, point cloud registration, and stitching.
Shadow Generation for Composite Image Using Diffusion Model
Qingyang Liu (Shanghai Jiao Tong University), Li Niu
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A shadow generation method based on diffusion models (SGDiffusion) is proposed, which can generate natural shadows for inserted foregrounds in synthetic images.
Shadow-Enlightened Image Outpainting
Hang Yu (Shanghai University), Jiayan Qiu (University of Leicester)
GenerationData SynthesisGraph Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes using shadow information for image outpainting.
Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now
Ayush Sarkar (University of Illinois Urbana-Champaign), Anand Bhattad (Toyota Technological Institute at Chicago)
GenerationAnomaly DetectionConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: The study generates geometric consistency in images by constructing a discriminator that detects differences between generated images and real images using three types of geometric features: shadows, perspective fields, and line segments.
Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification
Bin Yang (Wuhan University), Mang Ye (Wuhan University)
RecognitionRetrievalTransformerContrastive LearningImageMultimodality
🎯 What it does: A shallow-deep collaborative learning framework (SDCL) based on Transformer is proposed for unsupervised visible-infrared portrait recognition.
SHAP-EDITOR: Instruction-Guided Latent 3D Editing in Seconds
Minghao Chen (Visual Geometry Group, University of Oxford), Andrea Vedaldi (Visual Geometry Group, University of Oxford)
GenerationData SynthesisKnowledge DistillationDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: A real-time text-driven 3D editing framework SHAP-EDITOR based on a 3D latent space is proposed, capable of completing global and local edits within one second.
ShapeMatcher: Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation
Yan Di (Technical University of Munich), Shan Gao (Northwestern Polytechnical University)
SegmentationRetrievalContrastive LearningPoint CloudMesh
🎯 What it does: A self-supervised unified framework called ShapeMatcher has been designed and implemented to handle 3D objects with partial observations and arbitrary poses. It achieves fine-grained reconstruction results through four steps: shape regularization, semantic segmentation, retrieval, and deformation.
ShapeWalk: Compositional Shape Editing Through Language-Guided Chains
Habib Slim (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
TransformerLarge Language ModelAuto EncoderMesh
🎯 What it does: Constructed the ShapeWalk dataset and trained a shape autoencoder-based neural editor to achieve natural language-driven 3D shape editing.
Sharingan: A Transformer Architecture for Multi-Person Gaze Following
Samy Tafasca (Idiap Research Institute), Jean-Marc Odobez (Idiap Research Institute)
Object DetectionObject TrackingTransformerImageVideo
🎯 What it does: This paper proposes a Transformer-based multi-person eye tracking architecture called Sharingan, which can predict the gaze points of all individuals in an image at once.
Sheared Backpropagation for Fine-tuning Foundation Models
Zhiyuan Yu (University of Science and Technology of China), Dacheng Tao (Nanyang Technological University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes PreBackRazor, an activation pruning framework that combines gradient and activation joint sparsification for efficiently fine-tuning large-scale models on low-memory devices.
Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior
Fangfu Liu (Tsinghua University), Yueqi Duan (BAAI)
GenerationData SynthesisDiffusion modelScore-based ModelTextPoint CloudMesh
🎯 What it does: The Sherpa3D framework is proposed, which combines the coarse 3D prior generated by 3D diffusion with Score Distillation from 2D diffusion to generate high-fidelity, diverse, and multi-view consistent 3D models from text.
SHiNe: Semantic Hierarchy Nexus for Open-vocabulary Object Detection
Mingxuan Liu (University of Trento), Riccardo Volpi (NAVER LABS Europe)
Object DetectionVision Language ModelImage
🎯 What it does: This paper proposes a training-independent Semantic Hierarchical Association Classifier (SHiNe) that generates more robust text feature vectors by integrating hierarchical information of the target categories, thereby enhancing the performance of Open Vocabulary Object Detection (OvOD).
SHINOBI: Shape and Illumination using Neural Object Decomposition via BRDF Optimization In-the-wild
Andreas Engelhardt (University of Tübingen), Varun Jampani (Google Research)
Pose EstimationOptimizationNeural Radiance FieldImage
🎯 What it does: Jointly reconstruct the 3D shape, material (BRDF), and lighting information of each image from a collection of unlabeled natural environment images.
SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design
Seokju Yun (Machine Intelligence Laboratory, University of Seoul), Youngmin Ro (Machine Intelligence Laboratory, University of Seoul)
Object DetectionTransformerImage
🎯 What it does: A single-head visual Transformer SHViT is proposed, achieving low latency and high accuracy through a large stride patchify stem and a single-head attention module.
SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology
Saarthak Kapse (Stony Brook University), Prateek Prasanna (Stony Brook University)
ClassificationExplainability and InterpretabilityKnowledge DistillationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: A self-explanatory multi-instance learning framework named SI-MIL is proposed, which utilizes deep feature-guided interpretable branches to classify panoramic pathology slides and provide explanations at both regional and feature levels.
Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding
Chaolei Tan (Sun Yat-sen University), Jian-Fang Hu (Sun Yat-sen University)
Recurrent Neural NetworkTransformerVideoText
🎯 What it does: Proposes a weakly supervised video paragraph localization framework called SiamGTR, which can achieve multi-sentence temporal localization without timestamp annotations.
Sieve: Multimodal Dataset Pruning using Image Captioning Models
Anas Mahmoud (Meta), Ari S. Morcos (DatologyAI)
RetrievalData-Centric LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a multimodal dataset pruning method called Sieve, based on an image description generation model, to filter high-quality samples from noisy image-text pairs crawled from large-scale web pages.
SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction
Zechuan Zhang (Zhejiang University), Yi Yang (Zhejiang University)
GenerationData SynthesisPose EstimationTransformerDiffusion modelImageMesh
🎯 What it does: This paper proposes SIFU, which utilizes side-view conditional implicit functions and 3D consistent texture refinement to generate high-quality 3D models of clothed humans from a single image.
SIGNeRF: Scene Integrated Generation for Neural Radiance Fields
Jan-Niklas Dihlmann (University of Tuebingen), Hendrik Lensch (University of Tuebingen)
GenerationDiffusion modelNeural Radiance FieldImage
🎯 What it does: By first generating a reference table based on deep conditions, and then using this table for non-iterative image updates on the NeRF training set, rapid editing and object generation for existing NeRF scenes can be achieved.
SignGraph: A Sign Sequence is Worth Graphs of Nodes
Shiwei Gan (Nanjing University), Sanglu Lu (Nanjing University)
RecognitionGraph Neural NetworkVideoMultimodality
🎯 What it does: A continuous sign language recognition model called SignGraph based on graph convolutional networks has been constructed, utilizing local graphs (LSG) and temporal graphs (TSG) modules to capture cross-region and cross-frame features at the graph level, and learning sign language representations of different granularities at multiple scales.
SimAC: A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models
Feifei Wang (University of Science and Technology of China), Qidong Huang (University of Science and Technology of China)
GenerationSafty and PrivacyAdversarial AttackDiffusion modelImage
🎯 What it does: The SimAC method is proposed to suppress the model's personalized reproduction of user facial images by adding adversarial noise to the images in text-to-image diffusion models, thereby protecting user privacy.
SimDA: Simple Diffusion Adapter for Efficient Video Generation
Zhen Xing (Fudan University), Yu-Gang Jiang (Fudan University)
GenerationData SynthesisSuper ResolutionDiffusion modelVideo
🎯 What it does: Based on the Stable Diffusion pre-trained model, only 24M parameters (2% of 1.1B) are fine-tuned, achieving text-to-video (T2V) generation through lightweight spatial/time adapters and Latent-Shift Attention, and supporting one-click video editing and 1024×1024 super-resolution;
Simple Semantic-Aided Few-Shot Learning
Hai Zhang (Sichuan University), Zhenan He (Sichuan University)
ClassificationMeta LearningLarge Language ModelImage
🎯 What it does: This paper proposes a high-quality semantic description generation through an automated 'Semantic Evolution' process, and reconstructs class prototypes by fusing visual features with semantics using a simple two-layer Semantic Alignment Network, thereby enhancing few-shot classification performance.
Single Domain Generalization for Crowd Counting
Zhuoxuan Peng (Hong Kong University of Science and Technology), S.-H. Gary Chan (Hong Kong University of Science and Technology)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the MPCount model for single-domain generalization in crowd counting, addressing domain shift and label ambiguity issues.
Single Mesh Diffusion Models with Field Latents for Texture Generation
Thomas W. Mitchel (Google Research), Ameesh Makadia (Google Research)
GenerationData SynthesisCompressionDiffusion modelAuto EncoderMesh
🎯 What it does: This paper proposes an implicit diffusion model that operates directly on the surface of 3D shapes for synthesizing high-quality textures.
Single View Refractive Index Tomography with Neural Fields
Brandon Zhao (California Institute of Technology), Katherine L. Bouman (California Institute of Technology)
RestorationOptimizationNeural Radiance FieldImageComputed TomographyPhysics RelatedOrdinary Differential Equation
🎯 What it does: A method for single-view refractive index imaging using neural fields and differentiable curve ray tracing is proposed, which can recover a three-dimensional refractive index field from a single image.
Single-Model and Any-Modality for Video Object Tracking
Zongwei Wu (University of Wurzburg), Radu Timofte
Object TrackingTransformerVideoMultimodality
🎯 What it does: We propose Un-Track, a video object tracking model that can handle any single auxiliary modality such as RGB, depth, thermal, and event under the same set of parameters.