ICCV 2023 Papers — Page 19
IEEE/CVF International Conference on Computer Vision · 2156 papers
Stable Cluster Discrimination for Deep Clustering
Qi Qian (Alibaba Group)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes a one-stage deep clustering framework SeCu to address the convergence issue of cluster centers caused by traditional cross-entropy.
StableVideo: Text-driven Consistency-aware Diffusion Video Editing
Wenhao Chai (Zhejiang University), Yan Lu (Microsoft Research Asia)
GenerationData SynthesisDiffusion modelVideoStochastic Differential Equation
🎯 What it does: StableVideo is proposed, a text-driven diffusion model video editing framework that can edit the foreground and background of videos while maintaining geometric and temporal consistency.
StageInteractor: Query-based Object Detector with Cross-stage Interaction
Yao Teng (Nanjing University), Limin Wang (Nanjing University)
Object DetectionTransformerImage
🎯 What it does: This paper proposes a query-based object detector named StageInteractor, which integrates cross-stage label assignment and cross-stage dynamic filter reuse to enhance the model's ability for fine-grained discrimination and accelerate convergence.
Steered Diffusion: A Generalized Framework for Plug-and-Play Conditional Image Synthesis
Nithin Gopalakrishnan Nair (Johns Hopkins University), Tim K. Marks (Mitsubishi Electric Research Laboratories)
Image TranslationRestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: A Steered Diffusion framework based on an energy model is proposed, which guides an unconditional diffusion model during the inference phase using a pre-trained inverse mapping network (such as semantic maps, CLIP, VGGFace, etc.) without training a conditional model, achieving zero-shot conditional image generation and editing.
STEERER: Resolving Scale Variations for Counting and Localization via Selective Inheritance Learning
Tao Han (Shanghai Artificial Intelligence Laboratory), Wanli Ouyang (Hong Kong Polytechnic University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: The STEERER method is proposed to address the scale variation problem in object counting by selecting appropriate scales and inheriting scale-customized features.
StegaNeRF: Embedding Invisible Information within Neural Radiance Fields
Chenxin Li (Chinese University of Hong Kong), Zhangyang Wang (University of Texas at Austin)
RestorationOptimizationSafty and PrivacyNeural Radiance FieldImageMultimodalityAudio
🎯 What it does: This paper proposes a technique that can embed invisible information (such as images, text, audio, etc.) into the NeRF model, maintaining visual quality during the rendering of new perspectives while accurately recovering hidden information from the rendered images.
STEPs: Self-Supervised Key Step Extraction and Localization from Unlabeled Procedural Videos
Anshul Shah (Johns Hopkins University), Rama Chellappa (Johns Hopkins University)
TransformerContrastive LearningVideoMultimodality
🎯 What it does: Proposes an unsupervised method that uses self-supervised contrastive learning to automatically extract key steps from unlabeled instructional videos for AR-assisted training.
Stochastic Segmentation with Conditional Categorical Diffusion Models
Lukas Zbinden (University of Bern), Pablo Márquez-Neila (University of Bern)
SegmentationDiffusion modelImageComputed Tomography
🎯 What it does: A conditional classification diffusion model (CCDM) is proposed for semantic segmentation, which can generate diverse segmentation results that meet multimodal label distributions.
Story Visualization by Online Text Augmentation with Context Memory
Daechul Ahn (Yonsei University), Jonghyun Choi (Yonsei University)
GenerationData SynthesisTransformerImageTextMultimodality
🎯 What it does: A story visualization framework based on a bidirectional Transformer is proposed, utilizing a context memory module and online text augmentation technology to generate coherent and semantically consistent image sequences from paragraphs.
STPrivacy: Spatio-Temporal Privacy-Preserving Action Recognition
Ming Li (National University of Singapore), Shuicheng Yan (Sea AI Lab)
RecognitionSafty and PrivacyAdversarial AttackTransformerVideo
🎯 What it does: A video-level privacy-preserving action recognition framework called STPrivacy is proposed, which erases privacy information from videos while maintaining action recognition accuracy.
Strata-NeRF : Neural Radiance Fields for Stratified Scenes
Ankit Dhiman (Indian Institute of Science Bangalore), R Venkatesh Babu (Indian Institute of Science Bangalore)
GenerationData SynthesisNeural Radiance FieldAuto EncoderImage
🎯 What it does: A single neural radiance field named Strata-NeRF is proposed, capable of reconstructing a 3D representation of scenes with multi-level structures (layering) in a single training session, and achieving seamless switching between different levels of view synthesis.
Strip-MLP: Efficient Token Interaction for Vision MLP
Guiping Cao (Southern University of Science and Technology), Jianguo Zhang (Southern University of Science and Technology)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: The Strip-MLP model is proposed, which enhances token interaction through three mechanisms: the Strip MLP layer, CGSMM, and LSMM, thereby improving the performance of visual MLPs in image classification tasks.
Strivec: Sparse Tri-Vector Radiance Fields
Quankai Gao (University of Southern California), Zexiang Xu
GenerationData SynthesisOptimizationNeural Radiance FieldPoint Cloud
🎯 What it does: A sparse tri-vector radiance field (Strivec) is proposed as a neural representation of 3D scenes, utilizing local sparse 3D tensor grids and multi-scale CP decomposition to efficiently model geometry and appearance;
Structural Alignment for Network Pruning through Partial Regularization
Shangqian Gao (University of Pittsburgh), Heng Huang (University of Maryland)
CompressionOptimizationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A novel method for channel pruning is proposed through structural alignment and partial regularization.
Structure and Content-Guided Video Synthesis with Diffusion Models
Patrick Esser (Runway), Anastasis Germanidis (Runway)
GenerationData SynthesisDepth EstimationDiffusion modelImageVideo
🎯 What it does: This paper proposes a structure and content-guided video synthesis method based on a latent video diffusion model, which can edit video content according to text or image descriptions while maintaining the original video structure.
Structure Invariant Transformation for better Adversarial Transferability
Xiaosen Wang (Huawei Singularity Security Lab), Jianping Zhang (Chinese University of Hong Kong)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a Structure Invariant Transformation (SIT) and an attack method based on this transformation called Structure Invariant Attack (SIA). By locally applying various transformations on image blocks, it generates diverse inputs to enhance the transferability of adversarial samples.
Structure-Aware Surface Reconstruction via Primitive Assembly
Jingen Jiang, Dong-Ming Yan
GenerationOptimizationPoint CloudMesh
🎯 What it does: A structure-aware point cloud surface reconstruction method based on geometric primitive assembly is proposed, capable of generating manifold and closed 3D meshes.
Studying How to Efficiently and Effectively Guide Models with Explanations
Sukrut Rao (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
ClassificationObject DetectionOptimizationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: The study enhances the ability of multi-label classification models to focus on the features of target objects and improves generalization performance by utilizing interpretive information to guide the models.
StyleDiffusion: Controllable Disentangled Style Transfer via Diffusion Models
Zhizhong Wang (Zhejiang University), Wei Xing (Zhejiang University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: A controllable image style transfer framework called StyleDiffusion based on diffusion models is designed.
StyleDomain: Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation
Aibek Alanov (Higher School of Economics), Dmitry Vetrov (Higher School of Economics)
GenerationDomain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This study investigates the domain adaptation process of StyleGAN, systematically analyzing which network modules are sufficient for adaptation to domains of varying similarity, and proposes several lightweight parameterization schemes (StyleDomain direction, StyleSpaceSparse, Affine+ and AffineLight+) to achieve few-shot domain adaptation while supporting domain mixing and transfer.
StyleGANEX: StyleGAN-Based Manipulation Beyond Cropped Aligned Faces
Shuai Yang (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
Image TranslationGenerationSuper ResolutionGenerative Adversarial NetworkImageVideo
🎯 What it does: This paper proposes StyleGANEX, which modifies the first layer features to variable resolution using dilated convolutions in the shallow layers, expands the W+ space to a W+–F joint space, and retains the original parameters to handle untrimmed and unaligned panoramic portraits. It also constructs a corresponding encoder to achieve facial inversion and multi-task editing (attribute editing, super-resolution, sketch/mask to portrait translation, video cartoonization).
StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation
Yuhan Wang (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
GenerationData SynthesisGenerative Adversarial NetworkVideo
🎯 What it does: A non-autoregressive StyleInV framework is proposed, utilizing the inverse network of StyleGAN to generate motion latent variables through temporal style modulation, thereby achieving unconditional generation of long-sequence high-resolution videos.
StyleLipSync: Style-based Personalized Lip-sync Video Generation
Taekyung Ki (AITRICS), Dongchan Min (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkImageVideoAudio
🎯 What it does: This paper presents StyleLipSync, a personalized lip-sync video generation model based on StyleGAN, capable of generating high-quality, temporally consistent talking head videos from arbitrary audio and a single reference image.
StylerDALLE: Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model
Zipeng Xu (University of Trento), Nicu Sebe (University of Trento)
Image TranslationGenerationTransformerReinforcement LearningImageText
🎯 What it does: This paper proposes a language-guided style transfer method called Styler DALLE, which utilizes a large pre-trained vector quantization tokenizer and CLIP to generate stylized images that conform to text descriptions by translating discrete tokens of content images through non-autoregressive translation.
Subclass-balancing Contrastive Learning for Long-tailed Recognition
Chengkai Hou (Jilin University), Tianyi Zhou (University of Maryland)
ClassificationRecognitionObject DetectionContrastive LearningImage
🎯 What it does: A subclass balanced contrastive learning (SBCL) framework is designed to address the class imbalance problem in long-tail recognition.
SUMMIT: Source-Free Adaptation of Uni-Modal Models to Multi-Modal Targets
Cody Simons (University of California), Amit K. Roy-Chowdhury (University of California)
Domain AdaptationAutonomous DrivingImageMultimodalityPoint Cloud
🎯 What it does: Under the Source-Free Unsupervised Multimodal Adaptation (SUMMIT) framework, adaptive learning is performed on unlabelled paired multimodal data in the target domain using unimodal models independently trained in the source domain.
Supervised Homography Learning with Realistic Dataset Generation
Hai Jiang (Sichuan University), Shuaicheng Liu (University of Electronic Science and Technology of China)
Image TranslationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: An iterative framework is proposed that utilizes unlabeled real image pairs to generate training data that meets both labeling and realism criteria through the use of a dominant surface mask and estimated homography, and these data are used to supervise the training of the homography network.
SupFusion: Supervised LiDAR-Camera Fusion for 3D Object Detection
Yiran Qin (Chinese University of Hong Kong), Ruimao Zhang (NIO)
Object DetectionAutonomous DrivingKnowledge DistillationImagePoint Cloud
🎯 What it does: A supervised LiDAR-Camera 3D detection training strategy named SupFusion is proposed, which enhances detection accuracy through auxiliary feature supervision and a deep fusion module.
Surface Extraction from Neural Unsigned Distance Fields
Congyi Zhang (University of Hong Kong), Wenping Wang (Texas A&M University)
Point CloudMesh
🎯 What it does: The DualMesh-UDF method is proposed for extracting high-quality mesh surfaces from neural unsigned distance fields.
Surface Normal Clustering for Implicit Representation of Manhattan Scenes
Nikola Popovic (ETH Zurich), Luc Van Gool (ETH Zurich)
Depth EstimationRepresentation LearningNeural Radiance FieldImage
🎯 What it does: A self-supervised method is proposed to estimate Manhattan frames and constrain neural radiance fields, given the prior knowledge of the Manhattan scene but unknown Manhattan coordinate system.
SurfsUP: Learning Fluid Simulation for Novel Surfaces
Arjun Mani (Columbia University), Richard Zemel (Columbia University)
Graph Neural NetworkMesh
🎯 What it does: The SURFSUP framework is proposed, integrating surfaces in the form of implicit SDF into a graph neural network fluid simulator to learn and predict fluid-surface interactions.
SurroundOcc: Multi-camera 3D Occupancy Prediction for Autonomous Driving
Yi Wei (Tsinghua University), Jiwen Lu (Tsinghua University)
SegmentationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: A method called SurroundOcc is proposed, which utilizes multi-camera RGB images to predict dense 3D occupancy and generates dense occupancy labels through the stitching of multiple frames of LiDAR, Poisson reconstruction, and NN assignment.
SuS-X: Training-Free Name-Only Transfer of Vision-Language Models
Vishaal Udandarao (University of Cambridge), Samuel Albanie (University of Cambridge)
ClassificationRetrievalTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the SuS-X framework, which enables the transfer of visual-language models without training, relying solely on category names.
SVDFormer: Complementing Point Cloud via Self-view Augmentation and Self-structure Dual-generator
Zhe Zhu (Nanjing University of Aeronautics and Astronautics), Mingqiang Wei (Hong Kong Polytechnic University)
RestorationGenerationTransformerPoint Cloud
🎯 What it does: A point cloud completion network named SVDFormer is proposed, which utilizes self-projected multi-view depth maps to achieve global shape understanding and reconstructs fine-grained local details through a self-structured dual generator.
SVDiff: Compact Parameter Space for Diffusion Fine-Tuning
Ligong Han (Rutgers University), Feng Yang (Google Research)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Utilizing singular value decomposition (SVD) of diffusion models to fine-tune only the singular values (spectral shift) for model personalization and multi-agent generation, and proposing Cut-Mix-Unmix data augmentation.
SVQNet: Sparse Voxel-Adjacent Query Network for 4D Spatio-Temporal LiDAR Semantic Segmentation
Xuechao Chen (Tsinghua University), Lu Fang (Tsinghua University)
SegmentationAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkTransformerPoint Cloud
🎯 What it does: Proposes SVQNet, a Sparse Voxel Neighborhood Query Network for 4D LiDAR semantic segmentation.
SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications
Abdelrahman Shaker (Mohamed bin Zayed University of AI), Fahad Shahbaz Khan (Linköping University)
Object DetectionSegmentationComputational EfficiencyKnowledge DistillationTransformerImage
🎯 What it does: Designed the SwiftFormer architecture and proposed an efficient additive attention mechanism to achieve high accuracy and low latency for real-time visual tasks on mobile devices.
SwinLSTM: Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM
Song Tang (Hainan University), RongNian Tang (Hainan University)
Recurrent Neural NetworkTransformerVideoTime Series
🎯 What it does: A new temporal recursive unit called SwinLSTM is proposed, and a complete spatiotemporal prediction network is constructed based on this unit for future frame prediction.
SYENet: A Simple Yet Effective Network for Multiple Low-Level Vision Tasks with Real-Time Performance on Mobile Device
Weiran Gou (State Key Laboratory of Mobile Network and Mobile Multimedia Technology), Ke Xu (State Key Laboratory of Mobile Network and Mobile Multimedia Technology)
Super ResolutionConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented SYENet, a lightweight multi-task network with only 6K parameters, capable of real-time image signal processing (ISP), super-resolution (SR), and low-light enhancement (LLE) on mobile devices at 2K60FPS.
SynBody: Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling
Zhitao Yang (SenseTime Research), Lei Yang (Shanghai AI Laboratory)
Data SynthesisPose EstimationNeural Radiance FieldImageMultimodality
🎯 What it does: A large-scale synthetic dataset called SynBody has been constructed, containing 10,000 clothed SMPL-XL human models, 1.2 million high-quality images, and 2.7 million corresponding 3D annotations, supporting human pose and shape estimation as well as human NeRF research.
Synchronize Feature Extracting and Matching: A Single Branch Framework for 3D Object Tracking
Teli Ma (Hong Kong University of Science and Technology), Yong Liu (Zhejiang University)
Object TrackingAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes SyncTrack, a single-branch, single-stage 3D LiDAR single-object tracking framework that integrates feature extraction and matching.
Synthesizing Diverse Human Motions in 3D Indoor Scenes
Kaifeng Zhao (ETH Zurich), Siyu Tang (ETH Zurich)
GenerationPose EstimationRobotic IntelligenceReinforcement LearningMesh
🎯 What it does: A framework based on reinforcement learning is proposed, utilizing generative motion primitives and scene-aware strategies to automatically generate diverse human movements and human-object interactions in complex indoor 3D scenes.
Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models
Ziyi Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
ClassificationObject DetectionSegmentationGenerationTransformerImagePoint Cloud
🎯 What it does: A 3D-to-2D cross-modal generative pre-training method called TAP is proposed, which is applicable to any point cloud model. It pre-trains by using cross-attention on point cloud features to generate view images from different camera poses.
Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors
Zhentao Yu (Xiaobing AI), Baoyuan Wang (Xiaobing AI)
GenerationData SynthesisTransformerDiffusion modelGenerative Adversarial NetworkContrastive LearningVideoAudio
🎯 What it does: A method is proposed that can generate diverse and natural talking head videos using only audio and a single identity image, capable of producing non-lip-related facial movements such as posture, expressions, blinking, and gaze while maintaining audio-lip synchronization.
TALL: Thumbnail Layout for Deepfake Video Detection
Yuting Xu (Institute of Information Engineering, Chinese Academy of Sciences), Ran He (Institute of Automation, Chinese Academy of Sciences)
Anomaly DetectionComputational EfficiencyTransformerVideo
🎯 What it does: A Thumbnail Layout (TALL) strategy is proposed, which stitches video frames into thumbnails according to a fixed layout, capturing spatiotemporal inconsistencies without increasing computational load, thereby achieving deepfake video detection.
Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow
Federico Paredes-Vallés (Delft University of Technology), Guido C. H. E. de Croon (Sony Europe)
Autonomous DrivingRecurrent Neural NetworkContrastive LearningOptical FlowVideoSequential
🎯 What it does: This paper proposes a self-supervised event camera optical flow estimation framework that processes event streams sequentially through continuous recursive models.
Tangent Model Composition for Ensembling and Continual Fine-tuning
Tian Yu Liu (University of California), Stefano Soatto (University of California)
OptimizationComputational EfficiencySupervised Fine-TuningImage
🎯 What it does: The Tangent Model Composition (TMC) method is proposed, treating the fine-tuning of pre-trained models as tangent vectors, utilizing linear combinations to achieve model integration, continuous fine-tuning, and 'forgetting', thereby significantly reducing inference costs and storage requirements while maintaining high accuracy.
Tangent Sampson Error: Fast Approximate Two-view Reprojection Error for Central Camera Models
Mikhail Terekhov (ETH Zurich), Viktor Larsson (Lund University)
Pose EstimationOptimizationSimultaneous Localization and MappingImage
🎯 What it does: A novel two-view reprojection error metric called Tangent Sampson error is proposed, which is applicable to any central camera model and is mainly used for robust estimation and local refinement in two-view geometry.
TAPIR: Tracking Any Point with Per-Frame Initialization and Temporal Refinement
Carl Doersch (Google DeepMind), Andrew Zisserman (University College London)
Object TrackingConvolutional Neural NetworkVideo
🎯 What it does: A two-stage point tracking model called TAPIR is proposed for high-precision tracking of arbitrary query points in videos, capable of handling occlusions and long-term tracking.
TARGET: Federated Class-Continual Learning via Exemplar-Free Distillation
Jie Zhang (ETH Zurich), Lingjuan Lyu (Sony AI)
Data SynthesisFederated LearningKnowledge DistillationImage
🎯 What it does: In the framework of federated learning, the TARGET method is proposed to achieve federated class incremental learning through knowledge distillation without examples and generative data, thereby alleviating catastrophic forgetting.
Task Agnostic Restoration of Natural Video Dynamics
Muhammad Kashif Ali (Hanyang University), Tae Hyun Kim (Hanyang University)
RestorationConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes a task-agnostic framework for recovering the natural dynamics of videos by learning from inconsistent videos and utilizing consistent motion representations, thereby eliminating temporal flickering without relying on the original unprocessed video.
Task-aware Adaptive Learning for Cross-domain Few-shot Learning
Yurong Guo (Beijing University of Posts and Telecommunications), Zhanyu Ma (Beijing University of Posts and Telecommunications)
Domain AdaptationMeta LearningReinforcement LearningImage
🎯 What it does: A Task-aware Adaptive Network (TA-Net) is proposed for cross-domain few-shot learning, which enhances the model's adaptability by dynamically determining the number of task-specific adapters for each layer.
Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models
Sifan Long (Jilin University), Jingdong Wang (Baidu)
ClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a task-oriented multimodal mutual learning framework that combines Class-aware Text Prompting (CTP) and Text-guided Feature Tuning (TFT) to quickly adapt large visual-language models.
TaskExpert: Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts
Hanrong Ye (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
Object DetectionSegmentationTransformerMixture of ExpertsImage
🎯 What it does: The TaskExpert model is proposed, which achieves dynamic decoding of multi-task features through a mixture of multiple experts and dynamic gating.
Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation
Jianan Fan (University of Sydney), Weidong Cai (University of Sydney)
SegmentationDomain AdaptationOptimizationKnowledge DistillationImageBiomedical Data
🎯 What it does: An optimization trajectory distillation-based cross-domain adaptation framework is proposed, which can simultaneously address the issues of data distribution shift and label set inconsistency in medical images.
TCOVIS: Temporally Consistent Online Video Instance Segmentation
Junlong Li (Tsinghua University), Jiwen Lu (Tsinghua University)
Object DetectionSegmentationTransformerVideo
🎯 What it does: This paper proposes an online video instance segmentation method TCOVIS, which enhances temporal consistency through global instance allocation and a spatiotemporal enhancement module, enabling real-time inference by directly propagating queries between frames.
Teaching CLIP to Count to Ten
Roni Paiss (Google Research), Tali Dekel (Weizmann Institute of Science)
Object DetectionGenerationRetrievalTransformerVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: To address the poor object counting ability of visual language models like CLIP, we propose fine-tuning VLM with a counting contrastive loss and introduce a counting training set along with a new evaluation benchmark, CountBench, to improve the model's counting accuracy in zero-shot counting, retrieval, and text generation.
TeD-SPAD: Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection
Joseph Fioresi (University of Central Florida), Mubarak Shah (University of Central Florida)
Anomaly DetectionSafty and PrivacyContrastive LearningVideo
🎯 What it does: Designed and implemented the TeD-SPAD framework, achieving privacy protection in video anomaly detection through a self-supervised approach.
Tem-Adapter: Adapting Image-Text Pretraining for Video Question Answer
Guangyi Chen (Carnegie Mellon University), Yansong Tang (Tsinghua University)
Representation LearningTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Adapting the image-text pre-training model CLIP to the video question-answering task by adding a visual Temporal Aligner and a text Semantic Aligner.
Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction
Hatef Otroshi Shahreza (Idiap Research Institute), Sébastien Marcel (Idiap Research Institute)
RecognitionOptimizationAdversarial AttackNeural Radiance FieldGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the GaFaR method, which utilizes 3D facial reconstruction technology to reverse-generate high-quality 3D facial models from facial templates (i.e., feature vectors leaked from the FR system) and enhances the attack success rate through pose optimization.
Template-guided Hierarchical Feature Restoration for Anomaly Detection
Hewei Guo (Institute of Automation Chinese Academy of Sciences), Xinwen Hou (Institute of Automation Chinese Academy of Sciences)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A template-guided hierarchical feature recovery framework (THFR) is proposed for unsupervised anomaly detection and localization.
TEMPO: Efficient Multi-View Pose Estimation, Tracking, and Forecasting
Rohan Choudhury (Carnegie Mellon University), László A. Jeni (Carnegie Mellon University)
Object TrackingPose EstimationConvolutional Neural NetworkRecurrent Neural NetworkVideo
🎯 What it does: Proposes TEMPO, an efficient multi-view multi-person human pose estimation, tracking, and prediction framework;
Temporal Collection and Distribution for Referring Video Object Segmentation
Jiajin Tang (ShanghaiTech University), Sibei Yang (ShanghaiTech University)
Object DetectionSegmentationTransformerVideo
🎯 What it does: An end-to-end Temporal Query Collection and Distribution (TempCD) framework is proposed, which captures video-level reference semantics and frame-level object segmentation using global reference tokens and local object queries.
Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction
Zhuofan Zong (SenseTime Research), Yu Liu (SenseTime Research)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes a historical object prediction (HoP) auxiliary task, which predicts the 3D object at a given moment by generating pseudo BEV features from historical timestamps during the training phase.
Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold: Learning with Event-Driven Backpropagation
Wenjie Wei (University of Electronic Science and Technology of China), Hong Chen (Tsinghua University)
ClassificationSpiking Neural NetworkImage
🎯 What it does: This study investigates a hierarchical dynamic threshold mechanism and a direct learning algorithm to enhance the performance of TTFS-encoded deep spiking neural networks.
Test Time Adaptation for Blind Image Quality Assessment
Subhadeep Roy (Indian Institute of Science), Rajiv Soundararajan (Indian Institute of Science)
Domain AdaptationContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised adaptation framework for blind image quality assessment (IQA) during testing, utilizing two types of self-supervised auxiliary tasks: group contrastive loss and ranking loss, to adapt pre-trained models without accessing the source data.
Test-time Personalizable Forecasting of 3D Human Poses
Qiongjie Cui (Nanjing University of Science and Technology), Haofan Wang (Xiaohongshu Inc)
Pose EstimationDomain AdaptationKnowledge DistillationGenerative Adversarial NetworkTime Series
🎯 What it does: A helper-predictor structure is proposed, which achieves personalized 3D motion prediction for unseen target individuals through adaptive updates during testing.
Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra
Jonas Kulhanek, Torsten Sattler
Representation LearningNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes a neural radiance field representation based on sparse point clouds and adaptive tetrahedral meshes, called Tetra-NeRF. It utilizes Delaunay triangulation to obtain a set of tetrahedra and employs barycentric interpolation and a shallow MLP for volume rendering of point cloud features.
TexFusion: Synthesizing 3D Textures with Text-Guided Image Diffusion Models
Tianshi Cao (NVIDIA), Kangxue Yin (NVIDIA)
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: A large-scale text-guided image diffusion model is used to generate high-quality, globally consistent UV textures for a given 3D mesh.
Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models
Jaewoong Lee (KAIST), Sung Ju Hwang (KAIST)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A text-based sampling framework TCTS+FAS is proposed to improve text-conditioned image generation in token-based mask diffusion models.
Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization
Zhenhuan Liu (Institute of Computing Technology), Qingming Huang (University of Chinese Academy of Sciences)
GenerationDomain AdaptationGenerative Adversarial NetworkImageText
🎯 What it does: This paper proposes a text-driven domain adaptation method based on spectral consistency regularization, enabling pre-trained GANs to adapt to various target domains without image samples.
Text2Performer: Text-Driven Human Video Generation
Yuming Jiang (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: This paper proposes a text-based artificial intelligence framework called Text2Performer, which can generate high-quality, temporally consistent human videos based on input appearance and action descriptions, while maintaining identity features such as facial and clothing characteristics.
Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models
Lukas Höllein (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationData SynthesisDepth EstimationDiffusion modelMesh
🎯 What it does: Generate textured 3D meshes of complete room scale from text descriptions.
Text2Tex: Text-driven Texture Synthesis via Diffusion Models
Dave Zhenyu Chen (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: Using a pre-trained depth-aware diffusion model, driven by text prompts, high-quality textures are generated and updated for a given 3D mesh from multiple viewpoints in an iterative generate-refine process, dynamically partitioning view generation masks and automatically selecting the next best viewpoint for texture refinement.
Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators
Levon Khachatryan (Picsart AI Research), Humphrey Shi (Picsart AI Research)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: A zero-shot text-to-video generation method is proposed, utilizing a pre-trained text-to-image diffusion model (Stable Diffusion) and achieving video synthesis through a two-step lightweight modification.
TextManiA: Enriching Visual Feature by Text-driven Manifold Augmentation
Moon Ye-Bin (POSTECH), Tae-Hyun Oh (Yonsei University)
ClassificationObject DetectionTransformerLarge Language ModelImage
🎯 What it does: A text-based manifold augmentation method called TextManiA is proposed, which enhances the semantics of individual samples by projecting the text embedding difference vectors of attribute words into the visual feature space, achieving a densification of visual features.
TextPSG: Panoptic Scene Graph Generation from Textual Descriptions
Chengyang Zhao (Peking University), Chuang Gan (Massachusetts Institute of Technology IBM Watson AI Lab)
Object DetectionGenerationGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
🎯 What it does: A framework called TextPSG is proposed to generate Panoptic Scene Graphs (PSG) from pure text descriptions (image captions), addressing the weakly supervised scene graph generation task without the need for location priors, missing entity-region links, and unknown concept sets.
Texture Generation on 3D Meshes with Point-UV Diffusion
Xin Yu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: A Point-UV Diffusion framework based on a two-stage diffusion model is proposed for generating high-quality textures on 3D meshes of arbitrary topology.
Texture Learning Domain Randomization for Domain Generalized Segmentation
Sunghwan Kim (Agency for Defense Development), Hoseong Kim (Agency for Defense Development)
SegmentationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: The Texture Learning Domain Randomization (TLDR) framework is proposed, which utilizes texture learning to enhance the performance of domain generalization semantic segmentation models.
TF-ICON: Diffusion-Based Training-Free Cross-Domain Image Composition
Shilin Lu (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
Image TranslationGenerationDomain AdaptationDiffusion modelImageBenchmarkOrdinary Differential Equation
🎯 What it does: A training-free cross-domain image composition framework, TF-ICON, has been developed, which seamlessly integrates user-specified objects across various visual domains such as real images, oil paintings, sketches, and animations using a text-driven diffusion model.
The Devil is in the Crack Orientation: A New Perspective for Crack Detection
Zhuangzhuang Chen (Shenzhen University), Jianqiang Li (Shenzhen University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A crack detection method based on the direction of sub-cracks, called CrackDet, is proposed, treating cracks as a series of directed sub-cracks for detection.
The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior
Yilin Liu (University of North Carolina at Chapel Hill), Pew-Thian Yap (University of North Carolina at Chapel Hill)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper reveals through spectral analysis that the unlearned upsampling in Deep Image Prior is a key factor driving denoising, and based on this, proposes a strategy that only requires adjusting depth, width, and skip connections to automatically generate efficient denoising networks for each image.
The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining
Mannat Singh (Meta AI), Ishan Misra (Meta AI)
RecognitionRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: Proposes to add a self-supervised MAE pre-pretraining phase before large-scale weakly supervised pretraining to enhance the convergence and transfer performance of visual models.
The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation
Lingxiao Li (Columbia University), Shuhui Wang (Institute of Computing Technology Chinese Academy of Sciences)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: Proposes the Hyperbolic Attribute Editing (HAE) method, utilizing hyperbolic space to achieve hierarchical attribute editing for generating few-shot images.
The Making and Breaking of Camouflage
Hala Lamdouar (Shanghai Jiao Tong University), Andrew Zisserman (University of Oxford)
SegmentationGenerationData SynthesisConvolutional Neural NetworkTransformerGenerative Adversarial NetworkOptical FlowImageVideo
🎯 What it does: Three scoring functions for automatically evaluating the effectiveness of animal camouflage are proposed. These scoring functions are used to rank existing camouflage datasets, construct a generation pipeline capable of producing high-quality camouflage images and videos, and ultimately achieve optimal camouflage breaking performance based on synthetic data trained on the MoCA-Mask video dataset.
The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning
Virat Shejwalkar (University of Massachusetts Amherst), Amir Houmansadr (University of Massachusetts Amherst)
ClassificationAdversarial AttackImage
🎯 What it does: A backdoor injection attack based on unlabeled data is proposed, where a weak attacker can inject a strong backdoor into various semi-supervised learning (SSL) algorithms by contaminating only 0.2% of unlabeled data.
The Power of Sound (TPoS): Audio Reactive Video Generation with Stable Diffusion
Yujin Jeong, Jinkyu Kim
GenerationData SynthesisRecurrent Neural NetworkDiffusion modelVideoMultimodalityAudio
🎯 What it does: A Stable Diffusion-based audio-driven video generation framework called TPoS is designed, which can generate corresponding video frames in real-time based on the temporal semantics of audio.
The Stable Signature: Rooting Watermarks in Latent Diffusion Models
Pierre Fernandez (Meta AI Centre), Teddy Furon (Inria)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: Fine-tune the decoder of the Latent Diffusion Model (LDM) generator to embed an invisible watermark in all generated images, supporting detection and identification of the generation source;
The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation
Giacomo Zara (University of Trento), Elisa Ricci (Fondazione Bruno Kessler)
Domain AdaptationKnowledge DistillationTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: A source-agnostic video unsupervised domain adaptation method DALL-V is proposed, utilizing large language-vision models (such as CLIP) to assist the model in transferring from the source domain to the target domain.
The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data
Zixuan Zhu (Institute of Information Engineering, Chinese Academy of Sciences), Lihua Jing (Institute of Information Engineering, Chinese Academy of Sciences)
ClassificationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper studies a dual-network training framework called V&B, which uses a contaminated model to identify and filter poisoned samples, trains a clean model, and removes backdoors through semi-supervised suppression.
Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences
Felix Rydell (KTH Royal Institute of Technology), Angélica Torres (Centre de Recerca Matemàtica)
Depth EstimationOptimizationPoint Cloud
🎯 What it does: Research on 3D reconstruction under the point-line conjugate relationship, introducing anchored multi-view diversity and providing its Euclidean distance degree.
Thinking Image Color Aesthetics Assessment: Models, Datasets and Benchmarks
Shuai He (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
TransformerImageBenchmark
🎯 What it does: Proposes the task of image color aesthetic assessment, constructs the ICAA17K dataset, and introduces the Delegate Transformer baseline model.
TiDAL: Learning Training Dynamics for Active Learning
Seong Min Kye (Hyperconnect), Buru Chang (Sogang University)
ClassificationImage
🎯 What it does: This paper proposes and implements TiDAL, an active learning framework that utilizes training dynamics to predict the uncertainty of unlabeled samples.
TiDy-PSFs: Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions
Sachin Shah (University of Maryland), Christopher A. Metzler (University of Maryland)
Depth EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The paper proposes the use of time-averaged dynamic phase masks (TiDy-PSFs) to achieve multi-phase mask sequences, improving monocular depth estimation and panoramic imaging.
TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering
Yushi Hu (University of Washington), Noah A. Smith (Allen Institute for AI)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A no-reference evaluation metric TIFA based on visual question answering is proposed to measure the fidelity of text-to-image generation models.
TIJO: Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models
Indranil Sur (SRI International), Susmit Jha (SRI International)
Object DetectionOptimizationAdversarial AttackImageTextMultimodality
🎯 What it does: An offline defense method for multi-modal backdoor attacks, TIJO (Trigger Inversion using Joint Optimization), is proposed, which detects backdoor models by jointly optimizing to reverse the triggers of both image and text modalities.
Tiled Multiplane Images for Practical 3D Photography
Numair Khan (Meta), Douglas Lanman (Meta)
Depth EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A multi-layer image patch representation (TMPI) based on a single RGB image is proposed, utilizing adaptive depth planes to achieve lightweight 3D photography.
Time Does Tell: Self-Supervised Time-Tuning of Dense Image Representations
Mohammadreza Salehi (University of Amsterdam), Yuki M. Asano (University of Amsterdam)
SegmentationRepresentation LearningTransformerContrastive LearningOptical FlowImageVideo
🎯 What it does: A self-supervised temporal tuning method (Time-Tuning) is proposed, which enhances dense representations by performing temporal consistency clustering on pre-trained image models using unlabeled videos.
Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera
Urbano Miguel Nunes (Sorbonne University), Sio-Hoi Ieng (Sorbonne University)
Depth EstimationAutonomous DrivingOptimizationSimultaneous Localization and MappingOptical FlowTime Series
🎯 What it does: Proposes an incremental event processing method based on a single event camera, jointly estimating inverse depth (relative scale) and global motion, while maintaining a time-to-contact map (TTCM) in real-time, and providing optical flow estimation for each event.
Tiny Updater: Towards Efficient Neural Network-Driven Software Updating
Linfeng Zhang (Institute for Interdisciplinary Information Sciences, Tsinghua University), Kaisheng Ma (Institute for Interdisciplinary Information Sciences, Tsinghua University)
ClassificationComputational EfficiencyKnowledge DistillationImageVideo
🎯 What it does: The Tiny Updater method is proposed, which only downloads about 10%–20% of the parameters during network software updates, enhancing user experience.