CVPR 2023 Papers — Page 22
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
Token Boosting for Robust Self-Supervised Visual Transformer Pre-Training
Tianjiao Li (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
ClassificationRecognitionTransformerAuto EncoderImage
🎯 What it does: Proposes the Token Boosting Module (TBM), which enhances features of damaged inputs during self-supervised masked autoencoding pre-training in visual Transformers.
Token Contrast for Weakly-Supervised Semantic Segmentation
Lixiang Ru (Wuhan University), Bo Du (Wuhan University)
SegmentationTransformerContrastive LearningImage
🎯 What it does: By designing two contrastive modules, Patch Token Contrast (PTC) and Class Token Contrast (CTC), in the Vision Transformer, the issue of over-smoothing produced by ViT is addressed, thereby improving the weakly supervised semantic segmentation performance using only image-level labels.
Token Turing Machines
Michael S. Ryoo (Google Research), Anurag Arnab (Google Research)
RecognitionObject DetectionRobotic IntelligenceTransformerVideo
🎯 What it does: Proposes the Token Turing Machine (TTM), an autoregressive Transformer with external memory, designed for efficiently handling long-sequence visual understanding tasks;
TokenHPE: Learning Orientation Tokens for Efficient Head Pose Estimation via Transformers
Cheng Zhang (Central China Normal University), Youfu Li (City University of Hong Kong)
Pose EstimationTransformerImage
🎯 What it does: A TokenHPE method based on Transformer is proposed, which achieves head pose estimation by learning the relationships of key facial parts, showing excellent performance especially in extreme poses and occlusion scenarios.
Top-Down Visual Attention From Analysis by Synthesis
Baifeng Shi (University of California Berkeley), Xin Wang (Microsoft Corporation)
ClassificationSegmentationRetrievalTransformerImage
🎯 What it does: This paper proposes a top-level visual attention mechanism based on the analysis-synthesis (AbS) perspective and designs a controllable self-attention Transformer (AbSViT).
TopDiG: Class-Agnostic Topological Directional Graph Extraction From Remote Sensing Images
Bingnan Yang (Wuhan University), Xiangyun Hu (Wuhan University)
Object DetectionSegmentationTransformerImage
🎯 What it does: An end-to-end framework named TopDiG is proposed, capable of directly extracting multi-class (buildings, rivers, roads, etc.) topological directed graphs from remote sensing images.
TOPLight: Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition
Hao Yu (Nanjing University of Information Science and Technology), Wei Peng (Stanford University)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: A lightweight network called TOPLight is proposed, specifically designed for visible-infrared cross-modal recognition (person re-identification and face recognition).
TopNet: Transformer-Based Object Placement Network for Image Compositing
Sijie Zhu (University of Central Florida), Chen Chen (University of Central Florida)
GenerationData SynthesisTransformerContrastive LearningImage
🎯 What it does: This paper proposes a Transformer-based object placement network called TopNet, which can generate a 3D heatmap containing all candidate positions and scales in a single forward pass, thus enabling automated image synthesis for object placement.
Topology-Guided Multi-Class Cell Context Generation for Digital Pathology
Shahira Abousamra (Stony Brook University), Chao Chen (Stony Brook University)
SegmentationGenerationData SynthesisGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A generative model has been constructed that can generate multi-class cellular spatial layouts based on given pathological images, and the generated layouts are converted into H&E images for data augmentation.
ToThePoint: Efficient Contrastive Learning of 3D Point Clouds via Recycling
Xinglin Li (Hunan University), Di Wu (Hunan University)
ClassificationSegmentationRepresentation LearningContrastive LearningPoint Cloud
🎯 What it does: An efficient self-supervised contrastive learning framework called ToThePoint is proposed for pre-training 3D point cloud features.
Toward Accurate Post-Training Quantization for Image Super Resolution
Zhijun Tu (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
RestorationSuper ResolutionImage
🎯 What it does: This paper studies a post-training quantization method for single image super-resolution models and proposes a density-based dual cropping and pixel-aware calibration technique, allowing for the creation of a high-precision low-bit quantization model using only a small number of unlabeled calibration images.
Toward RAW Object Detection: A New Benchmark and a New Model
Ruikang Xu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
Object DetectionConvolutional Neural NetworkImageBenchmark
🎯 What it does: This paper presents the first large-scale 24-bit HDR RAW image dataset, ROD, and designs an end-to-end learnable image-level and pixel-level dynamic range adjustment network specifically for RAW sensor data, significantly improving object detection performance on RAW data.
Toward Stable, Interpretable, and Lightweight Hyperspectral Super-Resolution
Wen-jin Guo (Xidian University), Leyuan Fang (Hunan University)
RestorationSuper ResolutionOptimizationExplainability and InterpretabilityComputational EfficiencyAuto EncoderImage
🎯 What it does: A lightweight hyperspectral image super-resolution framework based on coordinated optimization is proposed, utilizing explicit degradation estimation and a sparse mixture prior autoencoder to recover high spatial resolution HSI.
Toward Verifiable and Reproducible Human Evaluation for Text-to-Image Generation
Mayu Otani (CyberAgent), Shin’ichi Satoh
GenerationData SynthesisDiffusion modelImageTextReview/Survey PaperBenchmark
🎯 What it does: This paper conducts a systematic investigation and summarizes the current state of human evaluation in the text-to-image generation field, finding a lack of standardization and poor reproducibility. Subsequently, a unified evaluation protocol based on absolute scoring is designed and validated, along with the public release of implementations, templates, and evaluation data.
Towards a Smaller Student: Capacity Dynamic Distillation for Efficient Image Retrieval
Yi Xie, Shengfeng He
RetrievalComputational EfficiencyKnowledge DistillationImage
🎯 What it does: Unable to access the paper content, further information is needed
Towards Accurate Image Coding: Improved Autoregressive Image Generation With Dynamic Vector Quantization
Mengqi Huang (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
GenerationCompressionTransformerAuto EncoderImage
🎯 What it does: A two-stage generative framework is constructed: first, a Dynamic Quantization VAE (DQ-VAE) allocates variable-length codes to different information density regions of the image, and then a DQ-Transformer generates the image in an autoregressive manner from coarse to fine.
Towards All-in-One Pre-Training via Maximizing Multi-Modal Mutual Information
Weijie Su (University of Science and Technology of China), Jifeng Dai (Tsinghua University)
Object DetectionSegmentationRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: A unified multimodal mutual information maximization framework M3I Pre-training is proposed, integrating supervised, weakly supervised, and self-supervised pre-training in a single-stage training;
Towards Artistic Image Aesthetics Assessment: A Large-Scale Dataset and a New Method
Ran Yi (Shanghai Jiao Tong University), Paul L. Rosin (Cardiff University)
ClassificationRecommendation SystemConvolutional Neural NetworkImage
🎯 What it does: This paper presents a large-scale aesthetic assessment dataset for artistic images, BAID (60,337 artworks, 360,000+ votes), and designs the Style-specific Art Assessment Network (SAAN) model for automatic evaluation of the aesthetic quality of artistic images.
Towards Benchmarking and Assessing Visual Naturalness of Physical World Adversarial Attacks
Simin Li (Beihang University), Xianglong Liu
Autonomous DrivingAdversarial AttackImageBenchmark
🎯 What it does: This study investigates the visual naturalness of physical world adversarial attacks and proposes an automatic evaluation method.
Towards Better Decision Forests: Forest Alternating Optimization
Miguel Á. Carreira-Perpiñán (University of California), Arman Zharmagambetov (University of California)
OptimizationTabularBenchmark
🎯 What it does: The Forest Alternating Optimization (FAO) algorithm is proposed, utilizing Tree Alternating Optimization (TAO) for global optimization across all forest parameters, replacing traditional greedy tree construction and ensemble methods.
Towards Better Gradient Consistency for Neural Signed Distance Functions via Level Set Alignment
Baorui Ma (Tsinghua University), Zhizhong Han (Wayne State University)
GenerationOptimizationNeural Radiance FieldPoint Cloud
🎯 What it does: A cascade alignment loss is proposed and implemented, achieving isosurface alignment without relying on annotated SDF by constraining the cosine distance between query points and their projected gradients on the zero isosurface, thereby enhancing the gradient consistency of neural SDF and improving 3D reconstruction of point clouds and multi-view images.
Towards Better Stability and Adaptability: Improve Online Self-Training for Model Adaptation in Semantic Segmentation
Dong Zhao (Xidian University), Licheng Jiao (Xidian University)
SegmentationDomain AdaptationImage
🎯 What it does: This paper proposes a source-agnostic semantic segmentation model adaptation method called DT-ST, which implements online self-training through dynamic teacher updates and training consistency resampling.
Towards Bridging the Performance Gaps of Joint Energy-Based Models
Xiulong Yang (Georgia State University), Shihao Ji (Georgia State University)
ClassificationGenerationImageStochastic Differential Equation
🎯 What it does: An improved Joint Energy-based Model (JEM) called SADA-JEM is proposed, which utilizes a single network to simultaneously perform image classification and generation.
Towards Building Self-Aware Object Detectors via Reliable Uncertainty Quantification and Calibration
Kemal Oksuz (Five AI Ltd), Puneet K. Dokania (Five AI Ltd)
Object DetectionDomain AdaptationImage
🎯 What it does: This paper proposes the Self-Aware Object Detection (SAOD) task, which requires detectors to have reliable uncertainty estimation, good calibration, and robustness to domain transfer in safety-critical scenarios.
Towards Compositional Adversarial Robustness: Generalizing Adversarial Training to Composite Semantic Perturbations
Lei Hsiung (National Tsing Hua University), Tsung-Yi Ho (National Tsing Hua University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A combination adversarial training (GAT) is proposed, which enhances the model's composite robustness by training on adversarial samples generated from various semantic perturbations and their combinations.
Towards Domain Generalization for Multi-View 3D Object Detection in Bird-Eye-View
Shuo Wang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
Object DetectionDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: A domain generalization framework for multi-view 3D object detection, DG-BEV, is proposed, which can maintain high detection accuracy in unseen domains.
Towards Effective Adversarial Textured 3D Meshes on Physical Face Recognition
Xiao Yang (Tsinghua University), Jun Zhu (Tsinghua University)
RecognitionAdversarial AttackImageMultimodalityMesh
🎯 What it does: This study designs and implements an adversarial texture mesh (AT3D) that can be 3D printed in real environments for physical black-box attacks on facial recognition systems while circumventing multimodal anti-spoofing mechanisms.
Towards Effective Visual Representations for Partial-Label Learning
Shiyu Xia (Southeast University), Xin Geng (Southeast University)
Representation LearningContrastive LearningImage
🎯 What it does: A partial label learning framework named PaPi is proposed, eliminating the contrastive learning module and using prototype alignment to guide the linear classifier's self-teaching, significantly improving representation learning and label disambiguation performance.
Towards Efficient Use of Multi-Scale Features in Transformer-Based Object Detectors
Gongjie Zhang (Nanyang Technological University), Shijian Lu (SenseTime Research)
Object DetectionPose EstimationComputational EfficiencyTransformerImage
🎯 What it does: A framework named Iterative Multi-Scale Feature Aggregation (IMFA) is proposed to improve the efficiency and effectiveness of Transformer-based object detectors in utilizing multi-scale features.
Towards End-to-End Generative Modeling of Long Videos With Memory-Efficient Bidirectional Transformers
Jaehoon Yoo (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisTransformerGenerative Adversarial NetworkVideo
🎯 What it does: This work proposes a Memory-efficient Bidirectional Transformer (MeBT) for end-to-end long video generation.
Towards Fast Adaptation of Pretrained Contrastive Models for Multi-Channel Video-Language Retrieval
Xudong Lin (Columbia University), Shih-Fu Chang (National University of Singapore)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This study investigates how to quickly adapt pre-trained contrastive learning models (multimodal and text) for multi-channel video-language retrieval tasks under limited data and computational resources.
Towards Flexible Multi-Modal Document Models
Naoto Inoue (CyberAgent), Kota Yamaguchi (CyberAgent)
GenerationData SynthesisTransformerSupervised Fine-TuningMultimodality
🎯 What it does: Proposes FlexDM, a unified multimodal document model capable of performing various design tasks (such as element filling, attribute prediction, layout generation, etc.) within the same framework.
Towards Generalisable Video Moment Retrieval: Visual-Dynamic Injection to Image-Text Pre-Training
Dezhao Luo (Queen Mary University of London), Yang Liu (Peking University)
RetrievalTransformerContrastive LearningVideoText
🎯 What it does: This paper proposes a Visual-Dynamic Injection (VDI) method to adapt image-text pre-trained models to video temporal retrieval tasks, enabling general retrieval of video segments.
Towards High-Quality and Efficient Video Super-Resolution via Spatial-Temporal Data Overfitting
Gen Li (Clemson University), Xiaolong Ma (Clemson University)
RestorationSuper ResolutionConvolutional Neural NetworkVideo
🎯 What it does: An adaptive video super-resolution method based on spatial-temporal information, STDO, is proposed. It segments the video into blocks and overfits each block, then utilizes data-aware joint training (JSTDO) to achieve a single model that covers the entire video.
Towards Modality-Agnostic Person Re-Identification With Descriptive Query
Cuiqun Chen (Wuhan University), Ding Jiang (Wuhan University)
RecognitionRetrievalTransformerContrastive LearningTextMultimodality
🎯 What it does: This paper studies a modality-agnostic person re-identification framework called UNIReID, which can simultaneously handle text, sketches, and joint queries of both, addressing the issue of modality uncertainty.
Towards Open-World Segmentation of Parts
Tai-Yu Pan (Ohio State University), Brian Price (Adobe Research)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes an Open Part Segmenter (OPS) method for open-world part segmentation, capable of accurately segmenting parts of unseen object categories.
Towards Practical Plug-and-Play Diffusion Models
Hyojun Go (Riiid AI Research), Seungtaek Choi (Riiid AI Research)
SegmentationGenerationKnowledge DistillationMixture of ExpertsDiffusion modelImage
🎯 What it does: The PPAP framework is proposed, utilizing multiple experts, parameter-efficient fine-tuning, and unsupervised knowledge transfer, allowing publicly available offline models to achieve conditional guidance in diffusion models in a plug-and-play manner.
Towards Professional Level Crowd Annotation of Expert Domain Data
Pei Wang (University of California San Diego), Nuno Vasconcelos (University of California San Diego)
ClassificationData-Centric LearningImage
🎯 What it does: A framework that integrates human filters into semi-supervised learning (SSL-HF) is proposed for quickly obtaining high-quality labels in expert domains (such as birds, fungi, etc.) through crowdsourcing.
Towards Realistic Long-Tailed Semi-Supervised Learning: Consistency Is All You Need
Tong Wei (Southeast University), Kai Gan (Southeast University)
ClassificationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An Adaptive Consistency Regularization (ACR) framework is proposed, utilizing a dual-branch network and dynamic logit adjustment to address the issue of unknown class distribution of unlabeled data in long-tail semi-supervised learning.
Towards Robust Tampered Text Detection in Document Image: New Dataset and New Solution
Chenfan Qu (South China University of Technology), Lianwen Jin (South China University of Technology)
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerImageText
🎯 What it does: A document image tampering text detection framework based on visual and frequency domain dual modalities is proposed—Document Tampering Detector (DTD). It introduces a technical solution utilizing Frequency Perception Head (FPH), Multi-view Iterative Decoder (MID), and Adaptive Compression Curriculum Learning (CLTD) to enhance detection accuracy and robustness. Additionally, a new tampered text dataset, DocTamper, consisting of 170,000 images, has been constructed.
Towards Scalable Neural Representation for Diverse Videos
Bo He (University of Maryland), Abhinav Shrivastava (University of Maryland)
RecognitionRestorationCompressionVideo
🎯 What it does: A scalable implicit neural representation D-NeRV is designed, capable of encoding diverse long videos in one go, achieving efficient compression and fast decoding.
Towards Stable Human Pose Estimation via Cross-View Fusion and Foot Stabilization
Li’an Zhuo, Liefeng Bo (Alibaba Group)
Pose EstimationTransformerImageVideo
🎯 What it does: This paper proposes a Cross-View Fusion (CVF) module and a Reversible Kinematic Topology Decoder (RKTD) to achieve more stable 3D human pose estimation from monocular images, and completes foot pose and foot-ground contact annotations for common datasets through a multi-view optimization method.
Towards Transferable Targeted Adversarial Examples
Zhibo Wang (Zhejiang University), Kui Ren (University of North Texas)
Adversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a new transferable target attack framework called TTAA, which utilizes generative adversarial networks combined with label and feature dual discriminators to generate highly transferable target adversarial samples.
Towards Trustable Skin Cancer Diagnosis via Rewriting Model's Decision
Siyuan Yan (Monash University), Zongyuan Ge (Monash University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a human-computer interactive framework that can automatically identify and correct biases in skin cancer diagnostic models caused by co-occurring factors such as image artifacts or backgrounds, making the model's inference more reliable.
Towards Unbiased Volume Rendering of Neural Implicit Surfaces With Geometry Priors
Yongqiang Zhang (NetEase Fuxi AI Lab), Changjie Fan (NetEase Fuxi AI Lab)
GenerationData SynthesisNeural Radiance FieldPoint Cloud
🎯 What it does: A method for converting unbiased perspective SDF to voxel rendering is proposed, combined with MVS point cloud supervision to achieve mask-free neural implicit surface reconstruction.
Towards Unified Scene Text Spotting Based on Sequence Generation
Taeho Kil (Naver Cloud), Daehee Kim (Naver Cloud)
RecognitionObject DetectionTransformerMixture of ExpertsImageText
🎯 What it does: A unified scene text detection and recognition model UNITS based on sequence generation is proposed, capable of handling four detection formats: center points, bounding boxes, quadrilaterals, and polygons within a single model, and achieving ultra-long sequence inference through starting point prompts.
Towards Universal Fake Image Detectors That Generalize Across Generative Models
Utkarsh Ojha (University of Wisconsin Madison), Yong Jae Lee (University of Wisconsin Madison)
ClassificationData SynthesisAnomaly DetectionTransformerVision Language ModelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper studies the problem of detecting fake images generated by generative models and finds that traditional binary classifiers have poor generalization ability across models.
Towards Unsupervised Object Detection From LiDAR Point Clouds
Lunjun Zhang (Waabi), Raquel Urtasun (Waabi)
Object DetectionAutonomous DrivingConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes an unsupervised LiDAR point cloud object detection method called OYSTER, which achieves zero-supervision object detection through near-distance point clustering, temporal consistency, adaptive expansion, and self-training.
Trace and Pace: Controllable Pedestrian Animation via Guided Trajectory Diffusion
Davis Rempe (NVIDIA), Or Litany
GenerationRobotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A controllable pedestrian trajectory generation method based on diffusion models, called TRACE, is proposed and combined with a physics-driven full-body controller PACER to form a closed-loop animation system.
TRACE: 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments
Yu Sun (Harbin Institute of Technology), Michael J. Black (Max Planck Institute for Intelligent Systems)
Object TrackingPose EstimationConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: We propose TRACE, a single-stage network based on 5D representation that can simultaneously regress the 3D poses, shapes, and global trajectories of multiple people under dynamic cameras.
Tracking Multiple Deformable Objects in Egocentric Videos
Mingzhen Huang (State University of New York at Buffalo), Siwei Lyu (Meta Reality Labs)
Object TrackingTransformerOptical FlowVideo
🎯 What it does: This paper presents DETracker, an end-to-end multi-object deformation tracking framework, and releases the first wearable camera collected DogThruGlasses dataset.
Tracking Through Containers and Occluders in the Wild
Basile Van Hoorick (Columbia University), Carl Vondrick (Columbia University)
Object TrackingSegmentationTransformerVideoBenchmark
🎯 What it does: This paper proposes the TCOW benchmark and corresponding models, studying the continuous tracking and segmentation of target objects in occlusion and containment scenarios.
Trade-Off Between Robustness and Accuracy of Vision Transformers
Yanxi Li (University of Sydney), Chang Xu (University of Sydney)
ClassificationAdversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: A model TORA-ViT is proposed, which inserts precision adapters and robustness adapters into the Vision Transformer, achieving a tunable trade-off between robustness and accuracy through attention-gated fusion.
Train-Once-for-All Personalization
Hong-You Chen (Ohio State University), Li Zhang (Ohio State University)
ClassificationRecognitionConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: This paper studies a framework called TAPER, which can quickly generate personalized models for multiple user tasks through task descriptions with a single training session, utilizing a baseline model and a mixer predictor to synthesize weights online.
Train/Test-Time Adaptation With Retrieval
Luca Zancato (AWS AI Labs), Stefano Soatto (AWS AI Labs)
RetrievalDomain AdaptationContrastive LearningImage
🎯 What it does: A method called TAR3 is proposed, which can adaptively pre-train models by retrieving external unlabeled samples during both training and inference.
Trainable Projected Gradient Method for Robust Fine-Tuning
Junjiao Tian (Georgia Institute of Technology), Zsolt Kira (Georgia Institute of Technology)
Domain AdaptationOptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: A trainable projection gradient method (TPGM) is proposed, which automatically learns the distance constraints between each layer and the pre-trained model during the fine-tuning process, thereby improving OOD robustness while maintaining ID performance.
Training Debiased Subnetworks With Contrastive Weight Pruning
Geon Yeong Park (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationContrastive LearningImage
🎯 What it does: A method for debiased subnetwork learning based on contrastive learning and weight pruning, called DCWP, is designed and proposed to mine unbiased subnetworks and fine-tune them using a small number of biased conflict samples on highly biased datasets.
Trajectory-Aware Body Interaction Transformer for Multi-Person Pose Forecasting
Xiaogang Peng (Hangzhou Dianzi University), Zizhao Wu (Hangzhou Dianzi University)
Pose EstimationTransformerTime SeriesSequential
🎯 What it does: A Transformer-based multi-person human pose prediction framework called TBIFormer is proposed, which can model fine-grained interactions and trajectory information of different body parts.
Transductive Few-Shot Learning With Prototype-Based Label Propagation by Iterative Graph Refinement
Hao Zhu (Australian National University), Piotr Koniusz (Australian National University)
ClassificationRecognitionGraph Neural NetworkImage
🎯 What it does: A transductive method called protoLP is proposed in few-shot learning, which integrates prototypes and label propagation, utilizing prototypes to construct a learnable graph and updating prototypes and label predictions in each iteration.
Transfer Knowledge From Head to Tail: Uncertainty Calibration Under Long-Tailed Distribution
Jiahao Chen (Renmin University of China), Bing Su (Renmin University of China)
ClassificationImage
🎯 What it does: A post-processing calibration method based on importance weights is proposed to address the issue of confidence miscalibration in models under long-tail distributions.
Transfer4D: A Framework for Frugal Motion Capture and Deformation Transfer
Shubh Maheshwari (TCS Research), Ramya Hebbalaguppe (Indian Institute of Technology Delhi)
GenerationPose EstimationVideo
🎯 What it does: This paper presents Transfer4D, a framework for unsupervised motion capture and animation transfer using a monocular depth camera, which can transfer motion from incomplete monocular depth videos to virtual characters of different shapes.
Transferable Adversarial Attacks on Vision Transformers With Token Gradient Regularization
Jianping Zhang (Chinese University of Hong Kong), Michael R. Lyu (Chinese University of Hong Kong)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the Token Gradient Regularization (TGR) method to enhance the transferable adversarial attack effectiveness of Vision Transformers (ViT).
TransFlow: Transformer As Flow Learner
Yawen Lu (Purdue University), Dongfang Liu (Rochester Institute of Technology)
TransformerOptical FlowImageVideo
🎯 What it does: A pure Transformer-based optical flow estimation model, TransFlow, is proposed, integrating spatial self-attention, cross-frame cross-attention, and multi-frame temporal association, equipped with a self-supervised pre-training module to achieve end-to-end optical flow estimation.
Transformer Scale Gate for Semantic Segmentation
Hengcan Shi (Monash University), Jianfei Cai (Monash University)
SegmentationTransformerImage
🎯 What it does: Designed and implemented the Transformer Scale Gate (TSG) module, which dynamically selects appropriate scale features using Vision Transformer attention information to enhance semantic segmentation performance.
Transformer-Based Learned Optimization
Erik Gärtner (Google Research), Cristian Sminchisescu (Google Research)
Pose EstimationOptimizationTransformerTabularTime Series
🎯 What it does: This paper proposes a Transformer-based learning optimizer called Optimus, which utilizes neural networks to learn each step's updates and precondition matrices, achieving more efficient iterative optimization.
Transformer-Based Unified Recognition of Two Hands Manipulating Objects
Hoseong Cho (Ulsan National Institute of Science and Technology), Seungryul Baek (Ulsan National Institute of Science and Technology)
RecognitionPose EstimationTransformerImageMesh
🎯 What it does: A unified framework based on Transformer (H2OTR) is proposed, capable of predicting the 3D poses of two hands, the object, the object type, and the categories of hand-object interaction actions all at once.
Transforming Radiance Field With Lipschitz Network for Photorealistic 3D Scene Stylization
Zicheng Zhang (University of Chinese Academy of Sciences), Ting Yao (HiDream.ai Inc.)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: Constructed the LipRF framework, utilizing the appearance representation of the pre-trained NeRF and the mapping of Lipschitz MLP to achieve cross-view consistency and lighting realism in 3D scene stylization.
TranSG: Transformer-Based Skeleton Graph Prototype Contrastive Learning With Structure-Trajectory Prompted Reconstruction for Person Re-Identification
Haocong Rao (Nanyang Technological University), Chunyan Miao (Nanyang Technological University)
RecognitionRetrievalGraph Neural NetworkTransformerContrastive LearningGraph
🎯 What it does: This paper proposes the TranSG framework for person re-identification (re-ID) based on 3D skeleton graphs.
Trap Attention: Monocular Depth Estimation With Manual Traps
Chao Ning (Northwestern Polytechnical University), Hongping Gan (Northwestern Polytechnical University)
Depth EstimationTransformerImage
🎯 What it does: This paper proposes a new Trap Attention mechanism that combines depthwise separable convolutions and manually defined trap functions, enabling global feature interaction while maintaining linear complexity, thus achieving monocular depth estimation.
Tree Instance Segmentation With Temporal Contour Graph
Adnan Firoze (Purdue University), Daniel Aliaga (Purdue University)
Object DetectionSegmentationGraph Neural NetworkOptical FlowImageVideo
🎯 What it does: This paper proposes a method for crown instance segmentation and counting based on RGB sequences collected by UAVs. It first performs over-segmentation on the sequences and constructs contour maps, then uses Graph Convolutional Networks (GCN) to learn contour features and complete crown merging, resulting in accurate instance masks and tree counts.
Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction
Yuanhui Huang (Tsinghua University), Jiwen Lu (Tsinghua University)
SegmentationAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: This paper proposes a Three-View (TPV) representation and the TPVFormer transformer model, which predicts 3D semantic occupancy volumes using only camera input combined with sparse LiDAR annotations, achieving complete spatial semantic occupancy prediction.
TriDet: Temporal Action Detection With Relative Boundary Modeling
Dingfeng Shi (Beihang University), Dacheng Tao (JD Explore Academy)
RecognitionObject DetectionConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a one-stage temporal action detection framework called TriDet, which combines the Trident-head and SGP layer to model the relative probability distribution of action boundaries and achieve scalable granularity perception in the feature pyramid.
TriVol: Point Cloud Rendering via Triple Volumes
Tao Hu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
GenerationData SynthesisNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes TriVol, a three-volume method for generating high-quality, hole-free rendered images from sparse point clouds.
TrojDiff: Trojan Attacks on Diffusion Models With Diverse Targets
Weixin Chen (University of Illinois Urbana-Champaign), Bo Li (University of Illinois Urbana-Champaign)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: The paper proposes a Trojan attack method for diffusion models called TrojDiff, which can inject triggers during training, allowing the model to generate preset attack targets when receiving trigger noise.
TrojViT: Trojan Insertion in Vision Transformers
Mengxin Zheng (Indiana University), Lei Jiang (Indiana University)
Knowledge DistillationAdversarial AttackTransformerImage
🎯 What it does: A backdoor attack method TrojViT for Vision Transformer (ViT) has been developed, utilizing patch-level triggers and bit flipping to implant malicious code.
TruFor: Leveraging All-Round Clues for Trustworthy Image Forgery Detection and Localization
Fabrizio Guillaro (University Federico II of Naples), Luisa Verdoliva (University Federico II of Naples)
Anomaly DetectionTransformerContrastive LearningImage
🎯 What it does: A framework named TruFor is proposed for trustworthy image forgery detection and localization.
TryOnDiffusion: A Tale of Two UNets
Luyang Zhu (University of Washington), Ira Kemelmacher-Shlizerman (Google Research)
GenerationData SynthesisPose EstimationDiffusion modelImage
🎯 What it does: A virtual clothing try-on method based on diffusion models is proposed, capable of synthesizing the target person and source clothing with detail retention at high resolution (1024×1024) while allowing significant changes in pose and body shape.
TTA-COPE: Test-Time Adaptation for Category-Level Object Pose Estimation
Taeyeop Lee (KAIST), Kuk-Jin Yoon (KAIST)
Pose EstimationDomain AdaptationImage
🎯 What it does: This paper proposes a test-time adaptive method for category-level object pose estimation (TTA-COPE), which can online update the model and improve pose prediction accuracy in the absence of source domain data and target labels.
Tunable Convolutions With Parametric Multi-Loss Optimization
Matteo Maggioni (Huawei), Aleš Leonardis (Huawei)
Image TranslationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A tunable convolution layer is proposed that can adjust the behavior of neural networks through interactive parameters during inference, combined with a parameterized multi-objective loss to achieve multi-objective control.
Turning a CLIP Model Into a Scene Text Detector
Wenwen Yu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Object DetectionDomain AdaptationTransformerPrompt EngineeringContrastive LearningImageText
🎯 What it does: Directly transform the pre-trained CLIP model into a scene text detector (TCM) without additional pre-training steps.
Turning Strengths Into Weaknesses: A Certified Robustness Inspired Attack Framework Against Graph Neural Networks
Binghui Wang (Illinois Institute of Technology), Yun Dong (Nanchang University)
Adversarial AttackGraph Neural NetworkGraph
🎯 What it does: A robust attack framework based on authentication has been designed, utilizing the authentication perturbation size of nodes to guide graph structure attacks, thereby enhancing the effectiveness of existing attack methods.
Twin Contrastive Learning With Noisy Labels
Zhizhong Huang (Fudan University), Hongming Shan (Fudan University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: Proposes the Twin Contrastive Learning (TCL) model, which combines contrastive learning with Gaussian Mixture Models (GMM) to learn robust representations from data with noisy labels and automatically correct the labels.
TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization
Ziquan Liu (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
ClassificationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A statistical information-based fine-tuning framework called TWINS is proposed to enhance the adversarial robustness and generalization performance of robust pre-trained models in downstream tasks.
Two-Shot Video Object Segmentation
Kun Yan (Peking University), Yan Lu (Microsoft Research Asia)
Object DetectionSegmentationVideo
🎯 What it does: This paper proposes a two-frame video object segmentation (Two-Shot VOS) training framework that uses annotations from only two frames of each video, leveraging pseudo-labels and semi-supervised learning to enhance model performance.
Two-Stage Co-Segmentation Network Based on Discriminative Representation for Recovering Human Mesh From Videos
Boyang Zhang (Ningxia University), Zhixiang Yuan (Ningxia University)
SegmentationPose EstimationConvolutional Neural NetworkRecurrent Neural NetworkVideoMesh
🎯 What it does: A Two-Stage Co-Segmentation Network is proposed, which constructs multi-layer discriminative features through spatial fine-grained segmentation and temporal fusion, ultimately using an SMPL regressor to recover the 3D mesh of humans in videos; a loss based on the area and perimeter of keypoint anchors is designed to constrain motion consistency.
Two-Stream Networks for Weakly-Supervised Temporal Action Localization With Semantic-Aware Mechanisms
Yu Wang (Ant Group), Hongbin Wang (Ant Group)
RecognitionObject DetectionConvolutional Neural NetworkVideo
🎯 What it does: A weakly supervised temporal action localization method is proposed, combining a dual-stream network with a semantic-aware mechanism.
Two-View Geometry Scoring Without Correspondences
Axel Barroso-Laguna (Niantic), Daniyar Turmukhambetov (Niantic)
Pose EstimationConvolutional Neural NetworkTransformerPoint Cloud
🎯 What it does: This paper proposes FSNet, a network that scores two-view geometric models (fundamental matrix or essential matrix) without using corresponding points, and combines it with traditional RANSAC loops.
Two-Way Multi-Label Loss
Takumi Kobayashi (National Institute of Advanced Industrial Science and Technology)
ClassificationObject DetectionSupervised Fine-TuningImage
🎯 What it does: A new multi-label loss function is proposed, which integrates the relative comparison idea of softmax with temperature scaling to explicitly enlarge the margin between positive and negative classes, and combines them into a bi-directional loss through both sample-wise and class-wise dimensions.
UDE: A Unified Driving Engine for Human Motion Generation
Zixiang Zhou, Baoyuan Wang
GenerationRetrievalTransformerDiffusion modelAuto EncoderTextMultimodalityAudio
🎯 What it does: A unified driving engine (UDE) is proposed to achieve human motion generation with dual-modal control of text and audio.
ULIP: Learning a Unified Representation of Language, Images, and Point Clouds for 3D Understanding
Le Xue (Salesforce Research), Silvio Savarese (Stanford University)
ClassificationRecognitionRepresentation LearningContrastive LearningMultimodalityPoint Cloud
🎯 What it does: This paper proposes the ULIP framework, which aligns 3D point cloud features using the pre-trained CLIP image-text feature space to construct a unified representation of language, images, and point clouds in three modalities.
Ultra-High Resolution Segmentation With Ultra-Rich Context: A Novel Benchmark
Deyi Ji (University of Science and Technology of China), Jieping Ye (Alibaba Group)
SegmentationConvolutional Neural NetworkImageBenchmark
🎯 What it does: A new benchmark dataset for UHR semantic segmentation, URUR, and an improved segmentation model, WSDNet, are proposed.
Ultrahigh Resolution Image/Video Matting With Spatio-Temporal Sparsity
Yanan Sun (Hong Kong University of Science and Technology), Yu-Wing Tai (Hong Kong University of Science and Technology)
SegmentationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImageVideo
🎯 What it does: This paper presents SparseMat, an efficient framework for high-resolution image/video matting;
UMat: Uncertainty-Aware Single Image High Resolution Material Capture
Carlos Rodriguez-Pardo (Universidad Rey Juan Carlos), Elena Garces (Universidad Rey Juan Carlos)
GenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a method that uses a single diffuse image captured by a flatbed scanner, combined with Generative Adversarial Networks (GAN), to predict the surface roughness, specularity, and normal map of textile materials, thereby achieving the digitalization of high-resolution SVBRDF.
Unbalanced Optimal Transport: A Unified Framework for Object Detection
Henri De Plaen (KU Leuven), Luc Van Gool (ETH Zurich)
Object DetectionOptimizationTransformerImage
🎯 What it does: This study investigates a unified matching framework based on Unbalanced Optimal Transport (UOT) to address the matching problem between predicted boxes and ground truth boxes in object detection training.
Unbiased Multiple Instance Learning for Weakly Supervised Video Anomaly Detection
Hui Lv (Nanjing University of Science and Technology), Hanwang Zhang (Nanyang Technological University)
Anomaly DetectionVideo
🎯 What it does: Proposed Unbiased Multiple Instance Learning (UMIL) for weakly supervised video anomaly detection.
Unbiased Scene Graph Generation in Videos
Sayak Nag (University of California), Amit K. Roy-Chowdhury (University of California)
Object DetectionGenerationTransformerContrastive LearningVideo
🎯 What it does: The TEMPURA framework is proposed for unbiased dynamic scene graph generation, addressing issues such as long-tail distribution, noise, and temporal inconsistency.
Uncertainty-Aware Optimal Transport for Semantically Coherent Out-of-Distribution Detection
Fan Lu (University of Science and Technology of China), Yang Cao (University of Science and Technology of China)
Anomaly DetectionOptimizationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper addresses the Semantically Coherent Out-of-Distribution (SCOOD) detection task and proposes an energy-based uncertainty-aware optimal transport (ET) scheme. By assigning semantically consistent labels to unlabeled data and combining a cross-set expansion strategy (Lrep) to enhance semantic discrimination, it achieves more accurate OOD detection.
Uncertainty-Aware Unsupervised Image Deblurring With Deep Residual Prior
Xiaole Tang (University of Electronic Science and Technology of China), Tieyong Zeng (Chinese University of Hong Kong)
RestorationImage
🎯 What it does: An unsupervised semi-blind image deblurring model is proposed, utilizing a dataset-independent deep residual prior (DRP) to model the residuals caused by blur kernel errors, and jointly recovering clear images with traditional handcrafted priors.
Uncertainty-Aware Vision-Based Metric Cross-View Geolocalization
Florian Fervers (Fraunhofer IOSB), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
Pose EstimationAutonomous DrivingTransformerImage
🎯 What it does: An end-to-end vision-based cross-view geolocation model has been developed, which can generate bird's-eye views from multiple camera perspectives and match them with aerial images, outputting a distribution of vehicle poses with uncertainty.
Uncovering the Disentanglement Capability in Text-to-Image Diffusion Models
Qiucheng Wu (University of California), Shiyu Chang (University of California)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This study investigates the intrinsic attribute decoupling capability of Stable Diffusion and proposes that optimizing only the mixed weights of two text embeddings can achieve various image editing tasks without fine-tuning the model.