CVPR 2023 Papers — Page 21
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2353 papers
SplineCam: Exact Visualization and Characterization of Deep Network Geometry and Decision Boundaries
Ahmed Imtiaz Humayun (Rice University), Richard G. Baraniuk
ClassificationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and implements SplineCam, a fully computable method for the geometry and decision boundaries of deep networks based on Continuous Piecewise Linear (CPWL) activation.
Spring: A High-Resolution High-Detail Dataset and Benchmark for Scene Flow, Optical Flow and Stereo
Lukas Mehl (University of Stuttgart), Andrés Bruhn (University of Stuttgart)
Data SynthesisDepth EstimationOptical FlowImageVideoBenchmark
🎯 What it does: This paper presents the Spring dataset and benchmark, providing 6000 frames of high-resolution (1920×1080) stereo images along with corresponding super-resolution (3840×2160) annotations for scene flow, optical flow, and disparity, aimed at evaluating dense matching performance in detail-rich scenes.
SQUID: Deep Feature In-Painting for Unsupervised Anomaly Detection
Tiange Xiang (University of Sydney), Zongwei Zhou (Johns Hopkins University)
Anomaly DetectionKnowledge DistillationTransformerImage
🎯 What it does: This paper proposes an unsupervised anomaly detection method called SQUID based on deep feature restoration for radiographic images.
sRGB Real Noise Synthesizing With Neighboring Correlation-Aware Noise Model
Zixuan Fu (Nanyang Technological University), Bihan Wen (Nanyang Technological University)
RestorationGenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A neighborhood correlation-aware (NeCA) sRGB noise synthesis framework is proposed, explicitly modeling the signal correlation and neighborhood correlation of noise, generating realistic camera noise directly in the sRGB domain.
Standing Between Past and Future: Spatio-Temporal Modeling for Multi-Camera 3D Multi-Object Tracking
Ziqi Pang (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
Object TrackingAutonomous DrivingTransformerImageVideo
🎯 What it does: An end-to-end multi-camera 3D multi-object tracking framework called PF-Track is designed, which integrates past and future spatiotemporal reasoning to achieve high-precision trajectory tracking.
STAR Loss: Reducing Semantic Ambiguity in Facial Landmark Detection
Zhenglin Zhou (Xiamen University), Rongrong Ji (Xiamen University)
Pose EstimationSupervised Fine-TuningImage
🎯 What it does: Proposes STAR loss, which utilizes the anisotropy of predicted heatmap distributions to alleviate the semantic ambiguity problem in facial keypoint annotation, and adaptively adjusts the error weights during training;
StarCraftImage: A Dataset for Prototyping Spatial Reasoning Methods for Multi-Agent Environments
Sean Kulinski (Purdue University), David I. Inouye (Purdue University)
RecognitionObject DetectionConvolutional Neural NetworkImageBenchmark
🎯 What it does: The StarCraftImage dataset is proposed, compressing 60k human StarCraft II match replays into 3.6M images, providing three easy-to-use representations: hyperspectral, RGB, and grayscale, facilitating rapid prototyping for multi-agent spatial reasoning methods.
Stare at What You See: Masked Image Modeling Without Reconstruction
Hongwei Xue (University of Science and Technology of China), Jiebo Luo (University of Rochester)
Object DetectionSegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: MaskAlign proposes a mask image modeling method that does not require reconstruction, directly aligning the visible patch features of the student model with the complete image features of the teacher model;
Starting From Non-Parametric Networks for 3D Point Cloud Analysis
Renrui Zhang (Chinese University of Hong Kong), Jianbo Shi (University of Pennsylvania)
ClassificationObject DetectionSegmentationPoint Cloud
🎯 What it does: A completely parameter-free, non-training parametric network called Point-NN is proposed for 3D point cloud classification, segmentation, and detection, and based on this, a parametric network Point-PN and an inference enhancement plugin are introduced.
STDLens: Model Hijacking-Resilient Federated Learning for Object Detection
Ka-Ho Chow (Georgia Institute of Technology), Yanzhao Wu (Georgia Institute of Technology)
Object DetectionFederated LearningConvolutional Neural NetworkImage
🎯 What it does: A three-layer defense framework called STDLens has been designed and implemented to identify and eliminate model hijacking attacks carried out through perception poisoning in federated learning.
SteerNeRF: Accelerating NeRF Rendering via Smooth Viewpoint Trajectory
Sicheng Li (Zhejiang University), Lu Yu (Zhejiang University)
Data SynthesisComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: A NeRF acceleration framework called SteerNeRF is proposed, which achieves real-time rendering by utilizing low-resolution volumetric rendering and high-resolution 2D neural rendering.
StepFormer: Self-Supervised Step Discovery and Localization in Instructional Videos
Nikita Dvornik (Samsung AI Centre Toronto), Allan D. Jepson (Samsung AI Centre Toronto)
TransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A self-supervised step discovery and localization framework called StepFormer is proposed, which can automatically discover and locate key steps in long, noisy instructional videos, achieving zero-shot multi-step localization without any manual annotations.
Stimulus Verification Is a Universal and Effective Sampler in Multi-Modal Human Trajectory Prediction
Jianhua Sun (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
Recurrent Neural NetworkGraph Neural NetworkMultimodalityTime Series
🎯 What it does: This paper proposes a general sampling method based on stimulus verification to improve the final prediction results of multimodal human trajectory prediction models.
Stitchable Neural Networks
Zizheng Pan (Monash University), Bohan Zhuang (Monash University)
Computational EfficiencyKnowledge DistillationNeural Architecture SearchConvolutional Neural NetworkTransformerImage
🎯 What it does: The Stitchable Neural Networks (SN-Net) framework is proposed, utilizing a family of pre-trained models as anchors, inserting 1×1 convolution stitching layers into the model to quickly generate sub-networks of various complexities and performances, achieving scalable deployment under dynamic resource constraints.
STMixer: A One-Stage Sparse Action Detector
Tao Wu (Nanjing University), Limin Wang (Nanjing University)
RecognitionObject DetectionTransformerVideo
🎯 What it does: Proposes STMixer, a single-stage sparse action detector;
STMT: A Spatial-Temporal Mesh Transformer for MoCap-Based Action Recognition
Xiaoyu Zhu (Carnegie Mellon University), Alexander G. Hauptmann
RecognitionPose EstimationTransformerMesh
🎯 What it does: Directly using the original mesh sequence for action recognition, avoiding the need to convert to a skeleton.
Streaming Video Model
Yucheng Zhao (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RecognitionObject TrackingTransformerVideo
🎯 What it does: A unified streaming video Transformer (S-ViT) architecture is proposed, which can be used for frame-level tasks (such as multi-object tracking) as well as sequence-level tasks (such as action recognition).
Structural Multiplane Image: Bridging Neural View Synthesis and 3D Reconstruction
Mingfang Zhang (University of Tokyo), Yan Lu (Microsoft Research Asia)
Data SynthesisDepth EstimationTransformerImage
🎯 What it does: A structured multi-plane image (SMPI) representation is proposed, which combines geometric planes with RGBα layers to generate high-quality new views from sparse images and achieve 3D plane reconstruction.
Structure Aggregation for Cross-Spectral Stereo Image Guided Denoising
Zehua Sheng (Zhejiang University), Huaqi Zhang (vivo Mobile Communication Company Ltd.)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a structure aggregation-based cross-spectral stereo image guided denoising network (SANet) that can recover details of noisy images from unaligned RGB-NIR image pairs.
Structured 3D Features for Reconstructing Controllable Avatars
Enric Corona (UPC), Cristian Sminchisescu (Google Research)
GenerationPose EstimationTransformerImageVideoPoint Cloud
🎯 What it does: Proposes Structured 3D Features (S3F) for animatable and relightable 3D human reconstruction from a single image, supporting multi-view aggregation and clothing editing.
Structured Kernel Estimation for Photon-Limited Deconvolution
Yash Sanghvi (Purdue University), Stanley H. Chan (Purdue University)
RestorationImage
🎯 What it does: A kernel estimation method based on low-dimensional motion trajectory key points is proposed for motion-blurred images under low-light photon-limited conditions, and blind deconvolution is completed through an iterative framework.
Structured Sparsity Learning for Efficient Video Super-Resolution
Bin Xia (Tsinghua University), Luc Van Gool (ETH Zurich)
RestorationSuper ResolutionCompressionComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: A structured sparse learning (SSL) framework is proposed, which achieves compression and acceleration of video super-resolution models through pruning residual blocks, recurrent networks, and pixel rearrangement.
StructVPR: Distill Structural Knowledge With Weighting Samples for Visual Place Recognition
Yanqing Shen (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
RetrievalKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes StructVPR, a knowledge distillation framework that utilizes semantic segmentation information to enhance the global features of RGB visual localization;
Style Projected Clustering for Domain Generalized Semantic Segmentation
Wei Huang (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
SegmentationDomain AdaptationAutonomous DrivingConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A method called 'Style Projected Clustering' is proposed, which projects the shallow features of unseen domain images using style basis vectors and achieves pixel-level category prediction through semantic basis vector clustering, thereby enhancing domain generalization performance in semantic segmentation.
StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot Learning
Yuqian Fu (Fudan University), Yu-Gang Jiang (Fudan University)
Domain AdaptationAdversarial AttackMeta LearningConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: A cross-domain few-shot learning method based on Style Adversarial Training (StyleAdv) is proposed, aimed at alleviating domain gaps by generating virtual and challenging samples in the style space.
StyleGAN Salon: Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer
Sasikarn Khwanmuang (Vistec), Supasorn Suwajanakorn (Vistec)
Image TranslationGenerationOptimizationGenerative Adversarial NetworkImage
🎯 What it does: By optimizing multi-view latent variables in the StyleGAN2 latent space, this method achieves pose-invariant hairstyle transfer, migrating the hairstyle from a reference image to the target person's face while keeping the facial identity and background unchanged.
StyleGene: Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis
Hao Li (Shenzhen University), Linlin Shen (Shenzhen University)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the StyleGene framework, which utilizes Regional Facial Genes (RFG) for kinship face synthesis, capable of generating high-fidelity and diverse offspring faces without the need for kinship-labeled data.
StyleIPSB: Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping
Diqiong Jiang (Zhejiang University), Min Tang (Zhejiang University)
Image TranslationGenerationGenerative Adversarial NetworkImageVideo
🎯 What it does: A semantic benchmark for identity preservation based on StyleGAN, called StyleIPSB, is proposed, and a three-stage high-fidelity face swapping framework is constructed to achieve precise fusion of the source face identity and target facial attributes.
StyleRes: Transforming the Residuals for Real Image Editing With StyleGAN
Hamza Pehlivan (Bilkent University), Aysegul Dundar (Bilkent University)
Image TranslationRestorationGenerationGenerative Adversarial NetworkImage
🎯 What it does: A single-stage image inversion framework called StyleRes is proposed, which learns residual features and performs transformations in a higher-order latent space, achieving high-fidelity reconstruction and high-quality attribute editing.
StyleRF: Zero-Shot 3D Style Transfer of Neural Radiance Fields
Kunhao Liu (Nanyang Technological University), Eric P. Xing
Image TranslationGenerationConvolutional Neural NetworkNeural Radiance FieldImage
🎯 What it does: Designed and implemented StyleRF, which can perform high-quality multi-view consistent style transfer on 3D scenes under zero-shot learning conditions.
StyLess: Boosting the Transferability of Adversarial Examples
Kaisheng Liang (Hong Kong Polytechnic University), Bin Xiao (Hong Kong Polytechnic University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: The StyLess attack method is proposed, which generates various stylized models by inserting instance normalization layers into the surrogate model, and jointly uses the gradients of these models to reduce reliance on non-robust style features, significantly enhancing the transferability of black-box attacks.
StyleSync: High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator
Jiazhi Guan (Tsinghua University), Jingdong Wang (Baidu Inc.)
GenerationData SynthesisGenerative Adversarial NetworkVideoAudio
🎯 What it does: The StyleSync framework is proposed to achieve high-fidelity and personalized lip synchronization.
SUDS: Scalable Urban Dynamic Scenes
Haithem Turki (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
Object DetectionObject TrackingSegmentationKnowledge DistillationRepresentation LearningNeural Radiance FieldOptical FlowVideoPoint Cloud
🎯 What it does: This paper studies a scalable neural radiance field representation method for dynamic urban scenes, SUDS, which can self-supervise the separation of static, dynamic, and distant objects across thousands of videos.
SunStage: Portrait Reconstruction and Relighting Using the Sun as a Light Stage
Yifan Wang (University of Washington), Xuaner Zhang (Adobe Inc.)
RestorationGenerationPose EstimationImageVideo
🎯 What it does: A lightweight lighting stage called SunStage is constructed using only a segment of outdoor video and sunlight, jointly reconstructing facial geometry, material, camera pose, and lighting information, which can be used for various post-processing tasks such as re-rendering, lighting editing, shadow softening, and view synthesis.
Super-CLEVR: A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning
Zhuowan Li (Johns Hopkins University), Alan L. Yuille
Object DetectionDomain AdaptationTransformerImageBenchmark
🎯 What it does: This paper proposes Super-CLEVR, a controllable virtual benchmark designed to separate and diagnose four domain shift factors in visual question answering (VQA) (visual complexity, question redundancy, concept distribution, and concept combination) and to evaluate model robustness.
Super-Resolution Neural Operator
Min Wei (Beijing University of Posts and Telecommunications), Xuesong Zhang (Beijing University of Posts and Telecommunications)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a continuous super-resolution framework based on neural operators (SRNO) that can recover high-resolution images from low-resolution images at arbitrary scales.
Superclass Learning With Representation Enhancement
Zeyu Gan (Renmin University of China), Cuiping Li (Renmin University of China)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A super-class learning framework named SCLRE is proposed, which enhances representations and reconstructs super-class decision boundaries using cross-instance attention.
SuperDisco: Super-Class Discovery Improves Visual Recognition for the Long-Tail
Yingjun Du (University of Amsterdam), Cees G. M. Snoek (University of Amsterdam)
ClassificationRecognitionMeta LearningGraph Neural NetworkImageBenchmark
🎯 What it does: By automatically learning multi-layer super-class graphs and combining them with message passing in graph neural networks, the original features in long-tail visual recognition are corrected and enhanced, thereby improving the recognition performance of tail classes.
Supervised Masked Knowledge Distillation for Few-Shot Transformers
Han Lin (Columbia University), Shih-Fu Chang (Columbia University)
Knowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: A supervised mask knowledge distillation (SMKD) framework is proposed to enhance model generalization performance in few-shot learning tasks using Vision Transformer.
SurfelNeRF: Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes
Yiming Gao (Tencent), Ying Shan (Tencent)
RestorationGenerationDepth EstimationRecurrent Neural NetworkNeural Radiance FieldPoint Cloud
🎯 What it does: For online realistic lighting reconstruction in indoor scenes, we propose using neural surface elements (Surfel) to construct a neural radiance field that can be fused and rendered online.
SVFormer: Semi-Supervised Video Transformer for Action Recognition
Zhen Xing (Fudan University), Yu-Gang Jiang (Fudan University)
RecognitionTransformerVideo
🎯 What it does: In the semi-supervised video action recognition task, the SVFormer framework is proposed, introducing Tube TokenMix and Temporal Warping Augmentation.
SVGformer: Representation Learning for Continuous Vector Graphics Using Transformers
Defu Cao (University of Southern California), Yan Liu (University of Southern California)
RetrievalRepresentation LearningTransformerImage
🎯 What it does: Proposes SVGformer, which directly encodes continuous SVG inputs using a Transformer to learn high-quality vector graphic representations.
SViTT: Temporal Learning of Sparse Video-Text Transformers
Yi Li (University of California San Diego), Nuno Vasconcelos (University of California San Diego)
RetrievalComputational EfficiencyTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes SViTT (Sparse Video-Text Transformer), a sparse video-text transformer that significantly reduces computational load and enhances the modeling capability of temporal relationships by achieving both edge sparsity and node sparsity in the attention map.
Swept-Angle Synthetic Wavelength Interferometry
Alankar Kotwal (Carnegie Mellon University), Ioannis Gkioulekas (Carnegie Mellon University)
Depth EstimationImage
🎯 What it does: A full-field micron-scale 3D sensing technology named swept-angle synthetic wavelength interferometry (SA-SWI) has been developed, and its high-precision depth reconstruction capability has been validated with experimental prototypes under complex materials (metals, translucent, scattering materials).
Switchable Representation Learning Framework With Self-Compatibility
Shengsen Wu (Peng Cheng Laboratory), Ling-Yu Duan (Peking University)
RecognitionRetrievalRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: A Switchable Representation Learning Framework with Self-compatibility (SFSC) is proposed, which obtains sub-models of different capacities through channel pruning on a full model. During the training process, uncertainty estimation and gradient projection are used to address gradient conflicts in multi-submodel collaborative optimization, achieving compatibility among multiple models.
Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion
Changfeng Ma (Nanjing University), Yanwen Guo (Nanjing University)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkPoint Cloud
🎯 What it does: Proposes an unsupervised symmetric shape-preserving autoencoder network (USSPA) for the completion of point clouds in real-world scenarios;
Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement
Siddarth Ravichandran (Samsung Research America), Hyun Jae Kang (Samsung Research America)
GenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkVideoAudio
🎯 What it does: An end-to-end real-time virtual face synthesis framework is proposed, which generates high-quality, well-synchronized virtual face videos using viseme audio features and facial keypoint mapping.
SynthVSR: Scaling Up Visual Speech Recognition With Synthetic Supervision
Xubo Liu (University of Surrey), Christian Fuegen (Meta AI)
RecognitionData SynthesisTransformerGenerative Adversarial NetworkVideoAudio
🎯 What it does: Proposes the SynthVSR framework, which uses a voice-driven mouth animation model to generate synthetic videos, and then conducts semi-supervised visual speech recognition training in conjunction with real videos.
System-Status-Aware Adaptive Network for Online Streaming Video Understanding
Lin Geng Foo (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
RecognitionPose EstimationComputational EfficiencyMeta LearningConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningVideo
🎯 What it does: An adaptive network with system state awareness (SAN) is proposed for achieving low-latency online video understanding on devices with fluctuating computational resources.
T-SEA: Transfer-Based Self-Ensemble Attack on Object Detection
Hao Huang (Peking University), Kevin Zhang (Peking University)
Object DetectionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A single-model self-ensemble adversarial patch attack (T-SEA) is designed to enhance the transferability of adversarial patches in multi-object detectors through strategies such as data self-ensemble, model ShakeDrop, and patch cutout.
Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation
Lingting Zhu (University of Hong Kong), Lequan Yu (University of Hong Kong)
GenerationData SynthesisPose EstimationTransformerDiffusion modelMultimodalityTime SeriesAudio
🎯 What it does: This paper proposes a framework for audio-driven co-speech gesture generation based on diffusion models, called DiffGesture, which can generate highly relevant and temporally coherent full-body pose sequences without using text or speaker identity.
Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation
Jun Nagata (DENSO IT LAB INC), Yusuke Sekikawa (DENSO IT LAB INC)
Autonomous DrivingOptical FlowImageVideo
🎯 What it does: This paper proposes an incremental full optical flow estimation method TEGBP based on event cameras, which infers complete optical flow from sparse normal flow.
TAPS3D: Text-Guided 3D Textured Shape Generation From Pseudo Supervision
Jiacheng Wei (Nanyang Technological University), Kim-Hui Yap (Nanyang Technological University)
GenerationData SynthesisVision Language ModelGenerative Adversarial NetworkPoint CloudMesh
🎯 What it does: A 3D texture shape generation model named TAPS3D was trained, capable of generating high-quality, controllable three-dimensional objects in one go based on given text prompts.
Target-Referenced Reactive Grasping for Dynamic Objects
Jirong Liu (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
Object DetectionObject TrackingRobotic IntelligenceContrastive LearningPoint Cloud
🎯 What it does: A dynamic object grasping method based on target referencing is proposed, utilizing a correspondence network and a memory-enhanced refinement network to track and maintain the temporal smoothness and semantic consistency of grasping poses across consecutive frames.
TarViS: A Unified Approach for Target-Based Video Segmentation
Ali Athar (RWTH Aachen University), Bastian Leibe (RWTH Aachen University)
Object TrackingSegmentationTransformerVideo
🎯 What it does: This paper proposes TarViS, a unified Transformer-based network that can perform video instance segmentation, video panoptic segmentation, video object segmentation, and point tracking within the same model.
Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning
Wenjin Wang (Zhejiang University), Yin Zhang (Zhejiang University)
Knowledge DistillationNeural Architecture SearchMixture of ExpertsImage
🎯 What it does: A lifelong learning framework called PAR is proposed, which can dynamically select parameter allocation or parameter regularization strategies based on the learning difficulty of tasks.
Task Residual for Tuning Vision-Language Models
Tao Yu (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes Task Residual Tuning (TaskRes), which adds learnable residual parameters to the CLIP pre-trained text classifier while keeping it unchanged to achieve efficient transfer learning.
Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification
Honglin Li (Zhejiang University), Lin Yang (Westlake University)
ClassificationDomain AdaptationComputational EfficiencyTransformerContrastive LearningImageBiomedical Data
🎯 What it does: A task-specific fine-tuning framework based on information bottleneck is proposed, which significantly improves weakly supervised whole slide image classification performance by utilizing multi-instance learning and SSL pre-trained features.
TBP-Former: Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving
Shaoheng Fang (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
Object DetectionSegmentationAutonomous DrivingTransformerImageVideo
🎯 What it does: This paper proposes a vision-based joint perception and prediction framework called TBP-Former, which utilizes a PoseSync BEV encoder and a spatial-temporal pyramid Transformer to detect, segment, and predict future trajectories of targets such as vehicles and pedestrians in synchronized BEV space.
Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models
Yushi Yao (Waymo), Gamaleldin F. Elsayed (Google)
ClassificationRetrievalContrastive LearningImage
🎯 What it does: The study utilizes human spatial attention labels as weak supervision, first training a teacher model on a small-scale attention dataset to generate pseudo-labels for ImageNet, and then incorporating these pseudo-labels as additional prediction targets into SimCLR contrastive learning to enhance the model's representation quality.
Teaching Matters: Investigating the Role of Supervision in Vision Transformers
Matthew Walmer (University of Maryland), Abhinav Shrivastava (University of Maryland)
ClassificationSegmentationRetrievalTransformerContrastive LearningImageVideo
🎯 What it does: This paper provides a comprehensive analysis of Vision Transformers trained with six different supervision methods (fully supervised, CLIP, DINO, MoCo, MAE, BEiT), covering attention patterns, feature representations, and downstream task performance.
Teaching Structured Vision & Language Concepts to Vision & Language Models
Sivan Doveh (IBM Research), Leonid Karlinsky (MIT-IBM Watson AI Lab)
ClassificationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningTextMultimodality
🎯 What it does: By processing the existing visual-text pair data, structured visual-language concept (SVLC) positive and negative samples are generated using methods such as rules, LLM, and analogy, supplemented by additional loss training to enhance the VL model's understanding of object attributes, relationships, and states of SVLC.
Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction
Ryo Kawahara (Kyushu Institute of Technology), Shohei Nobuhara (Kyoto University)
Depth EstimationComputational EfficiencyImage
🎯 What it does: This study proposes a 'Teleidoscopic Imaging System' that combines a single-focus lens and a kaleidoscope mirror to achieve multi-view close-range imaging and 3D reconstruction of micro-scale objects.
Tell Me What Happened: Unifying Text-Guided Video Completion via Multimodal Masked Video Generation
Tsu-Jui Fu (University of California Santa Barbara), Sean Bell (Meta)
GenerationData SynthesisTransformerGenerative Adversarial NetworkContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes the Text-Guided Video Completion (TVC) task, which can generate complete video sequences based on given starting frames, ending frames, or frames from both ends, along with natural language descriptions.
Temporal Attention Unit: Towards Efficient Spatiotemporal Predictive Learning
Cheng Tan (Westlake University), Stan Z. Li (Westlake University)
Computational EfficiencyConvolutional Neural NetworkVideo
🎯 What it does: In the video prediction task, a Temporal Attention Unit (TAU) is proposed as a parallelized temporal module, and differential divergence regularization is designed to enhance cross-frame difference learning.
Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving
Lucas Nunes (University of Bonn), Cyrill Stachniss (University of Bonn)
Object DetectionSegmentationAutonomous DrivingRepresentation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: This paper proposes a method for self-supervised representation learning that utilizes vehicle motion to obtain LiDAR views of the same object at different times, training the network to learn time-consistent and robust point cloud features against object dynamics.
Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields
Sungheon Park (Samsung Advanced Institute of Technology), Nahyup Kang (Samsung Advanced Institute of Technology)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldVideo
🎯 What it does: A dynamic NeRF training framework based on temporal interpolation is proposed, which includes two implementations: neural networks and hash grids.
Temporally Consistent Online Depth Estimation Using Point-Based Fusion
Numair Khan (Meta), Lei Xiao (Meta)
Depth EstimationOptical FlowVideoPoint Cloud
🎯 What it does: This paper proposes an online temporally consistent video depth estimation framework that achieves real-time consistent depth maps through a global point cloud-driven three-stage fusion.
TempSAL - Uncovering Temporal Information for Deep Saliency Prediction
Bahar Aydemir (École Polytechnique Fédérale de Lausanne), Sabine Süsstrunk (École Polytechnique Fédérale de Lausanne)
SegmentationGenerationConvolutional Neural NetworkImageTime Series
🎯 What it does: This paper proposes a deep model called TempSAL that can simultaneously predict attention heatmaps for each second time segment and the final attention heatmap for the entire image;
TensoIR: Tensorial Inverse Rendering
Haian Jin (Zhejiang University), Hao Su (UC San Diego)
GenerationOptimizationComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: TensoIR is proposed, utilizing low-rank tensor decomposition of neural field to jointly estimate scene geometry, material, and environmental lighting under unknown lighting conditions in multi-view images, achieving high-quality inverse rendering and relighting.
Tensor4D: Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering
Ruizhi Shao (Tsinghua University), Yebin Liu (Tsinghua University)
RestorationGenerationData SynthesisComputational EfficiencyNeural Radiance FieldImageVideo
🎯 What it does: This paper proposes Tensor4D, an efficient 4D spatio-temporal decomposition method for reconstructing and rendering dynamic scenes from sparse views or monocular videos.
TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation
Devavrat Tomar (EPFL), Jean-Philippe Thiran (EPFL)
Domain AdaptationKnowledge DistillationContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A self-learning method called TeSLA is proposed, which utilizes automatic adversarial augmentation for online adaptation of pre-trained models on unlabeled streaming test data.
Test of Time: Instilling Video-Language Models With a Sense of Time
Piyush Bagad (University of Amsterdam), Cees G. M. Snoek
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: Detect and enhance the existing video-language model's understanding of temporal order, proposing the TACT temporal adaptation method implemented on VideoCLIP.
Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection
Olga Veksler (University of Waterloo)
Object DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: For weakly supervised salient object detection, a method for adaptive fine-tuning of a single image during testing is proposed, utilizing regularization loss to achieve more accurate segmentation results without pixel-level labels.
TexPose: Neural Texture Learning for Self-Supervised 6D Object Pose Estimation
Hanzhi Chen (Technical University of Munich), Benjamin Busam (3Dwe.ai)
Pose EstimationNeural Radiance FieldImage
🎯 What it does: This paper proposes a method for self-supervised 6D pose estimation through neural texture learning.
Text With Knowledge Graph Augmented Transformer for Video Captioning
Xin Gu (University of Chinese Academy of Sciences), Longyin Wen (ByteDance Inc.)
GenerationTransformerVideoText
🎯 What it does: This paper proposes TextKG, a dual-stream Transformer model enhanced by knowledge graphs, aimed at generating more accurate video subtitles.
Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation
Xiwen Wei (South China University of Technology), Hau San Wong (City University of Hong Kong)
Image TranslationGenerationGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised multi-attribute image editing framework based on CLIP, named TUSLT, which can modify multiple attributes simultaneously in a single latent space transformation step and improve editing accuracy through a self-trained auxiliary attribute classifier.
Text-Visual Prompting for Efficient 2D Temporal Video Grounding
Yimeng Zhang (Michigan State University), Ke Ding (Intel)
RecognitionOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerPrompt EngineeringVideoTextMultimodality
🎯 What it does: A text-visual prompting framework (TVP) is proposed to enhance the performance of the time video grounding (TVG) model that uses only 2D sparse visual features, achieving end-to-end trainability.
Text2Scene: Text-Driven Indoor Scene Stylization With Part-Aware Details
Inwoo Hwang (Seoul National University), Young Min Kim (Seoul National University)
GenerationData SynthesisImageRetrieval-Augmented Generation
🎯 What it does: An end-to-end process named Text2Scene is proposed, which generates fine-grained, semantically bounded realistic textures for multiple objects in indoor 3D scenes using reference images and text guidance.
Texts as Images in Prompt Tuning for Multi-Label Image Recognition
Zixian Guo (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
ClassificationRecognitionPrompt EngineeringVision Language ModelImageText
🎯 What it does: This paper proposes treating text descriptions as images for prompt tuning (TaI prompting) and introduces dual-granularity prompt tuning (TaI-DPT) in multi-label image recognition to simultaneously leverage global and local features.
Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection
Huajun Zhou (Sun Yat-sen University), Xiaohua Xie (Sun Yat-sen University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkContrastive LearningImageVideoMultimodality
🎯 What it does: A framework for unsupervised salient object detection has been designed and implemented, which generates pseudo-labels through self-supervised activation maps and continuously optimizes label quality using these two novel mechanisms, ultimately training a high-quality salient detector.
The Best Defense Is a Good Offense: Adversarial Augmentation Against Adversarial Attacks
Iuri Frosio (NVIDIA), Jan Kautz (NVIDIA)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A new framework A5 (Adversarial Augmentation to Defend Against Adversarial Attacks) is proposed, which is the first certified preventive defense method against adversarial attacks, aimed at ensuring that any attack will fail by constructing defensive perturbations.
The Dark Side of Dynamic Routing Neural Networks: Towards Efficiency Backdoor Injection
Simin Chen (University of Texas at Dallas), Wei Yang (University of Texas at Dallas)
Computational EfficiencyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper studies and implements an attack method called EfficFrog that can implant an 'efficiency backdoor' in dynamic neural networks. The attacker can cause the victim model to consume more computational resources when triggered by specific inputs using a small amount of training data, leading to a decrease in system availability.
The Devil Is in the Points: Weakly Semi-Supervised Instance Segmentation via Point-Guided Mask Representation
Beomyoung Kim (NAVER Cloud), Sung Ju Hwang (NAVER AI Lab)
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Weakly Supervised Instance Segmentation (WSSIS) framework that utilizes point labels with just one pixel as a weak supervision source and refines coarse masks using MaskRefineNet.
The Dialog Must Go On: Improving Visual Dialog via Generative Self-Training
Gi-Cheon Kang (Seoul National University), Byoung-Tak Zhang (Seoul National University)
GenerationData SynthesisRetrievalTransformerVision Language ModelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: Proposes Generative Self-Training (GST) to expand the VisDial training set by generating multi-turn visual dialogue data on unlabeled images from the web.
The Differentiable Lens: Compound Lens Search Over Glass Surfaces and Materials for Object Detection
Geoffroi Côté (Université Laval), Felix Heide (Princeton University)
Object DetectionAutonomous DrivingOptimizationImage
🎯 What it does: This paper proposes a differentiable spherical compound lens simulation model and achieves lens design for object detection tasks by jointly optimizing lens parameters (curvature, spacing, glass materials) with the gradients of a downstream automotive object detection network.
The Enemy of My Enemy Is My Friend: Exploring Inverse Adversaries for Improving Adversarial Training
Junhao Dong (Sun Yat-Sen University), Xiaohua Xie (Sun Yat-Sen University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposes an Inverse Adversarial Training (IAT) framework that utilizes 'inverse adversarial samples' to guide the model in learning decision regions with higher credibility;
The ObjectFolder Benchmark: Multisensory Learning With Neural and Real Objects
Ruohan Gao (Stanford University), Jiajun Wu (Stanford University)
RecognitionRetrievalGenerative Adversarial NetworkImageVideoMultimodalityBenchmarkAudio
🎯 What it does: A benchmark suite called OBJECTFOLDER BENCHMARK has been constructed, which includes 10 multi-sensory object-centered learning tasks, and a multi-sensory dataset containing 100 real household items (visual, acoustic, tactile) has been released.
The Resource Problem of Using Linear Layer Leakage Attack in Federated Learning
Joshua C. Zhao (Purdue University), Saurabh Bagchi (Purdue University)
Federated LearningSafty and PrivacyComputational EfficiencyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: The study investigates the resource overhead of linear layer leakage attacks under secure aggregation and proposes a MANDRAKE attack that utilizes sparse convolutional layers to significantly reduce model size and computation time.
The Treasure Beneath Multiple Annotations: An Uncertainty-Aware Edge Detector
Caixia Zhou (Beijing Jiaotong University), Haibin Ling (Stony Brook University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A framework for uncertainty-aware edge detection (UAED) is proposed, which models the multi-label annotation space as a learnable Gaussian distribution, using variance to reflect annotation uncertainty and adaptively weighting the training process;
The Wisdom of Crowds: Temporal Progressive Attention for Early Action Prediction
Alexandros Stergiou (Vrije University of Brussels), Dima Damen (University of Bristol)
RecognitionTransformerVideo
🎯 What it does: Designed and implemented the TemPr model, which utilizes a multi-scale evolutionary attention tower for early action prediction on partially observed videos.
Therbligs in Action: Video Understanding Through Motion Primitives
Eadom Dessalene (University of Maryland), Yiannis Aloimonos (University of Maryland)
RecognitionSegmentationRecurrent Neural NetworkTransformerSupervised Fine-TuningVideo
🎯 What it does: In this paper, we propose a hierarchical rule reasoning framework based on Therblig for action understanding in videos.
Thermal Spread Functions (TSF): Physics-Guided Material Classification
Aniket Dashpute (Rice University), Oliver Cossairt (University of Arizona)
ClassificationRecognitionImagePhysics Related
🎯 What it does: Using low-power laser heating and recording the thermal diffusion process of objects with a thermal camera, the thermal diffusion function (TSF) is extracted. By solving the inverse heat conduction equation, the thermal diffusivity and emissivity are obtained, which are then used as features for material classification, ultimately achieving an accuracy of about 86% in multi-class recognition.
Think Twice Before Driving: Towards Scalable Decoders for End-to-End Autonomous Driving
Xiaosong Jia (Shanghai Jiao Tong University), Hongyang Li (Shanghai AI Laboratory)
Autonomous DrivingKnowledge DistillationRecurrent Neural NetworkTransformerReinforcement LearningMultimodalityPoint Cloud
🎯 What it does: This paper proposes the ThinkTwice end-to-end autonomous driving framework, which implements a scalable decision network through a multi-layer coarse-to-fine prediction decoder (Look, Predict, Refine).
Three Guidelines You Should Know for Universally Slimmable Self-Supervised Learning
Yun-Hao Cao (Nanjing University), Shuchang Zhou (MEGVII Technology)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A self-supervised model method US3L is proposed, which can be used at any width with a one-time training.
TimeBalance: Temporally-Invariant and Temporally-Distinctive Video Representations for Semi-Supervised Action Recognition
Ishan Rajendrakumar Dave (University of Central Florida), Mubarak Shah (University of Central Florida)
RecognitionKnowledge DistillationRepresentation LearningContrastive LearningVideo
🎯 What it does: Proposes the TimeBalance framework, which uses time-invariant and time-discriminative video representations obtained from self-supervised learning as two teacher models for semi-supervised action recognition.
TINC: Tree-Structured Implicit Neural Compression
Runzhao Yang (Tsinghua University)
CompressionBiomedical DataComputed Tomography
🎯 What it does: Proposes the Tree-structured Implicit Neural Compression (TINC) framework, which first divides large-scale data into blocks and uses MLP for local INR representation, then shares parameters in a tree hierarchical structure to enhance compression quality.
TinyMIM: An Empirical Study of Distilling MIM Pre-Trained Models
Sucheng Ren (Microsoft Research Asia), Han Hu (Microsoft Research Asia)
ClassificationSegmentationKnowledge DistillationTransformerImage
🎯 What it does: Knowledge is transferred from a large-scale Masked Image Modeling (MIM) pre-trained model to a small Vision Transformer (ViT) through knowledge distillation, thereby enhancing the performance of the small model.
TIPI: Test Time Adaptation With Transformation Invariance
A. Tuan Nguyen (University of Oxford), Philip H.S. Torr (University of Oxford)
ClassificationDomain AdaptationAdversarial AttackImage
🎯 What it does: During the inference phase, online adaptation for the pre-trained model is performed, proposing an unsupervised input transformation invariance regularization objective to prevent model collapse under small batch sizes.
TMO: Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering
Jaehoon Choi (NAVER LABS), Donghwan Lee (NAVER LABS)
Object DetectionDepth EstimationOptimizationNeural Radiance FieldSimultaneous Localization and MappingImageVideoPoint CloudMesh
🎯 What it does: This paper proposes a complete texture mesh reconstruction framework based on smartphones, utilizing mobile sensors to collect images, LiDAR depth, and pose, achieving high-quality textured meshes through RGB-D assisted SfM, neural implicit surface reconstruction, and differentiable rendering.