IEEE/CVF International Conference on Computer Vision Β· 743 papers
Spatial-Aware Token for Weakly Supervised Object Localization
Pingyu Wu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
CodeObject DetectionSegmentationTransformerImage
π― What it does: This paper proposes a weakly supervised object localization method based on Transformer, utilizing Spatial-Aware Tokens (SAT) to directly generate localization maps through spatial query attention, avoiding optimization conflicts between classification and localization tasks.
Spatially and Spectrally Consistent Deep Functional Maps
Mingze Sun (Tsinghua University), Ruqi Huang (Tsinghua University)
CodeDiffusion modelPoint CloudMesh
π― What it does: Proposes a two-branch unsupervised deep feature mapping (DFM) framework that integrates spectral domain and spatial domain cyclic consistency to enhance non-rigid shape correspondence.
π― What it does: A unified Spatio-Temporal Prompt Network (STPN) is proposed, achieving robust video feature extraction by injecting dynamic video prompts at the front end of the Transformer, eliminating complex backend integration modules.
Spectrum-guided Multi-granularity Referring Video Object Segmentation
Bo Miao (University of Western Australia), Ajmal Mian (University of Western Australia)
CodeObject DetectionSegmentationTransformerVideo
π― What it does: A spectrum-guided multi-granularity R-VOS framework SgMg is proposed to address the feature drift problem and achieve single-frame/multi-frame/multi-object segmentation.
π― What it does: A Speech2Lip framework is proposed, which generates high-fidelity, synchronized speaker videos from just a few minutes of video by separating speech-sensitive and speech-insensitive motions, utilizing implicit models, geometric mapping, hybrid networks, and contrastive synchronization loss.
π― What it does: A large-scale multi-object tracking dataset, SportsMOT, covering basketball, volleyball, and soccer scenarios has been constructed, and the MixSort framework has been proposed to enhance tracking performance.
π― What it does: This paper systematically discovers and verifies harmful interference features on ImageNet using class-level neural PCA methods and visualization techniques, and proposes the SpuFix correction scheme, which reduces the dependency of any ImageNet classifier on these interference features without the need for additional annotations or retraining.
π― What it does: The SQAD dataset is proposed, which conducts six quality assessments (resolution, color accuracy, noise, dynamic range, PSF, and aliasing) based on laboratory measurements for 29 smartphone cameras, and trains deep models to achieve automatic camera quality assessment and device identification.
π― What it does: A simple yet powerful baseline (SSB) is proposed in open-set semi-supervised learning, achieving improvements in both in-class classification and outlier detection through three main strategies: high-confidence pseudo-labeling, nonlinear feature projection heads, and pseudo-negative sample mining.
Sabbir Ahmed (Binghamton University), Adnan Siraj Rakin (North Carolina State University)
CodeDomain AdaptationAdversarial AttackImage
π― What it does: This study investigates the risks of backdoor attacks in source-free domain adaptation (SFDA) and proposes a secure training scheme called SSDA.
π― 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.
π― 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.
π― 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.
π― 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;
π― 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.
Studying How to Efficiently and Effectively Guide Models with Explanations
Sukrut Rao (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
CodeClassificationObject DetectionOptimizationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
π― 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.
π― 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.
π― 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.
π― 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.
π― 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.
π― 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.
π― 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.
π― 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)
CodeClassificationRetrievalTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
π― 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)
CodeRestorationGenerationTransformerPoint Cloud
π― 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.
SwinLSTM: Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM
Song Tang (Hainan University), RongNian Tang (Hainan University)
CodeRecurrent Neural NetworkTransformerVideoTime Series
π― 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)
π― 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.
π― 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.
Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation
Jianan Fan (University of Sydney), Weidong Cai (University of Sydney)
CodeSegmentationDomain AdaptationOptimizationKnowledge DistillationImageBiomedical Data
π― 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)
CodeObject DetectionSegmentationTransformerVideo
π― 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.
CodeRepresentation LearningTransformerVision Language ModelVideoTextMultimodality
π― 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.
Test Time Adaptation for Blind Image Quality Assessment
Subhadeep Roy (Indian Institute of Science), Rajiv Soundararajan (Indian Institute of Science)
CodeDomain AdaptationContrastive LearningImage
π― 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.
π― 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.
π― 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.
π― 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.
π― 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 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)
CodeRestorationConvolutional Neural NetworkImage
π― 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.
π― What it does: Proposes the Hyperbolic Attribute Editing (HAE) method, utilizing hyperbolic space to achieve hierarchical attribute editing for generating few-shot images.
The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation
Giacomo Zara (University of Trento), Elisa Ricci (Fondazione Bruno Kessler)
CodeDomain AdaptationKnowledge DistillationTransformerLarge Language ModelVision Language ModelVideo
π― 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)
π― 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.
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)
CodeTransformerImageBenchmark
π― 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)
CodeClassificationImage
π― What it does: This paper proposes and implements TiDAL, an active learning framework that utilizes training dynamics to predict the uncertainty of unlabeled samples.
π― 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.
π― 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)
CodeDepth EstimationAutonomous DrivingOptimizationSimultaneous Localization and MappingOptical FlowTime Series
π― 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.
Too Large; Data Reduction for Vision-Language Pre-Training
Alex Jinpeng Wang (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
CodeRetrievalCompressionTransformerVision Language ModelImageTextMultimodality
π― What it does: Proposes the TL;DR algorithm, which aligns codebook quantization with the generation of new captions, selects representative samples, and compresses large-scale VLP datasets.
Toward Multi-Granularity Decision-Making: Explicit Visual Reasoning with Hierarchical Knowledge
Yifeng Zhang (University of Minnesota), Qi Zhao (University of Minnesota)
CodeObject DetectionExplainability and InterpretabilityGraph Neural NetworkMultimodality
π― What it does: This paper proposes a Hierarchical Concept Graph (HCG) and a Hierarchical Concept Neural Module Network (HCNMN) to achieve explicit reasoning and explanation of multi-granularity knowledge in visual question answering.
Towards Attack-tolerant Federated Learning via Critical Parameter Analysis
Sungwon Han (KAIST), Meeyoung Cha (KAIST)
CodeFederated LearningAdversarial AttackImage
π― What it does: This paper proposes a new federated learning defense strategy called FedCPA, which uses key parameter analysis to assess the normality of model updates, thereby defending against poisoning attacks from malicious clients.
Towards Authentic Face Restoration with Iterative Diffusion Models and Beyond
Yang Zhao (Google), Matthias Grundmann (Google)
CodeRestorationGenerationDiffusion modelImage
π― What it does: A real face restoration system based on an iterative diffusion model (IDM) is proposed, which automatically cleans training data through external iterative learning.
π― What it does: A novel uncertainty robustness enhancement framework for unsupervised domain adaptation (UDA) called DDAR is proposed to enhance the model's robustness against common noise and distortions (RaCC).
π― What it does: A deep unified framework for depth-aware panoramic segmentation is proposed, achieving both instance-level semantic segmentation and monocular depth estimation.
π― What it does: An effective instance identification contrastive loss (EIDCo) specifically designed for domain adaptation is proposed, which learns unlabeled target domain features through low-confidence samples, class relationship enhanced features, and target-dominant cross-domain Mixup.
π― What it does: A modular code library was constructed, unified training standards were established, and an error diagnosis tool was proposed to systematically evaluate 3D detection based on images.
π― What it does: A geospatial foundation model (GFM) based on continuous pre-training is proposed, which combines ImageNet-22k weights with self-supervised MIM objectives for multi-task training.
π― What it does: A unified representation learning framework for unsupervised visible-infrared person recognition is proposed, addressing the differences at the camera and modality levels;
π― What it does: A text-guided 3D face generation model TG-3DFace has been developed, which can generate high-quality, multi-view consistent 3D faces and textures based solely on text-2D face image data.
π― What it does: This study investigates the issue of client internal heterogeneity in federated learning and proposes the FedIns algorithm, which reduces communication costs and improves model accuracy through instance adaptive reasoning and SSF pooling.
Towards Memory- and Time-Efficient Backpropagation for Training Spiking Neural Networks
Qingyan Meng (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)
CodeSpiking Neural NetworkImage
π― What it does: Proposes the Spatial Learning Through Time (SLTT) method, which improves the time and memory overhead of training SNNs using traditional BPTT+SG;
π― What it does: A geometry-based quadratic rolling shutter motion solver (QRS) and a 3D video structure for rolling shutter correction in extreme occlusion scenarios, RSA-Net2, are proposed, capable of achieving high-quality global shutter image reconstruction under complex nonlinear motion and occlusion conditions.
Towards Open-Vocabulary Video Instance Segmentation
Haochen Wang (University of Amsterdam), Efstratios Gavves (University of Amsterdam)
CodeObject TrackingSegmentationTransformerVideo
π― What it does: This paper proposes the Open-Vocabulary Video Instance Segmentation (OV-VIS) task, which aims to achieve segmentation, tracking, and classification of any category in videos simultaneously.
π― What it does: A real scene flash group super-resolution dataset (RealBSR) and the FBAnet model are proposed, utilizing isomorphic alignment, federated similarity fusion, and Transformer decoding to achieve multi-frame super-resolution.
Towards Realistic Evaluation of Industrial Continual Learning Scenarios with an Emphasis on Energy Consumption and Computational Footprint
Vivek Chavan (Fraunhofer Institute for Production Systems and Design Technology), Clemens Briese (Fraunhofer Institute for Production Systems and Design Technology)
π― What it does: A realistic evaluation of Class Incremental Learning (CIL) in industrial scenarios is conducted, proposing the RECIL framework and introducing the InVar-100 industrial object dataset, focusing on three-dimensional metrics: accuracy, energy consumption, and computational overhead.
π― What it does: This paper introduces two regularization verification mechanisms, self-sanity and cross-sanity, to constrain the training of existing deep image registration models, thereby reducing folding transformations, enhancing inverse consistency and discriminative ability, and providing a theoretical error upper bound.
π― What it does: A framework for adversarial sample tracing is proposed under the buyer-seller setting, divided into two phases: model separation and source tracking. It utilizes a parallel structure and a VAE-trained tracer to achieve model differentiation, and locates the source of the attack through the logit difference output by the tracer.
π― What it does: This paper proposes TrajectoryFormer, which utilizes predicted trajectory hypotheses and multiple candidates from detection boxes to perform transformer-style association and refinement of 3D trajectories, achieving robust LiDAR 3D multi-object tracking.
TrajPAC: Towards Robustness Verification of Pedestrian Trajectory Prediction Models
Liang Zhang (Institute of Software), Lijun Zhang (Institute of Software)
CodeAutonomous DrivingExplainability and InterpretabilityAdversarial AttackReinforcement LearningVideo
π― What it does: Proposed and implemented the TRAJPAC framework to validate the robustness of pedestrian trajectory prediction models, providing formal definitions of label robustness and pure robustness, and using the PAC method to complete robustness assessment and interpretability analysis of black-box models.
π― What it does: This paper transforms the pseudo-label generation task in weakly supervised semantic segmentation into an image matting problem and proposes the Mat-Label framework.
π― What it does: A completely data-independent universal adversarial perturbation generation method TRM-UAP is designed, capable of quickly generating attack perturbations on any CNN.
π― What it does: This paper proposes an Ethnic Quality Bias Mitigation framework (EQBM) based on curriculum-style domain adaptation, which adjusts the facial image quality assessment in the target domain by mapping the original regression targets to Likert quantized probabilities and incorporating a difficulty scheduler.
π― What it does: A self-supervised video representation learning method is proposed, which achieves contrastive learning of motion dynamics by inserting synthetic motion trajectories (tubelets) into the video.
π― What it does: The paper proposes a new feature distribution-based linear classifier called Moment Probing (MP), and designs a Partial Shared Recalibration Module (PSRP) based on it for efficient fine-tuning of pre-trained models.
π― What it does: This paper proposes a two-in-one network called TiO-Depth that can simultaneously perform monocular and binocular self-supervised depth estimation, and further enhances the prediction accuracy of both tasks through multi-stage joint training.
U-RED: Unsupervised 3D Shape Retrieval and Deformation for Partial Point Clouds
Yan Di (Technical University of Munich), Federico Tombari (Technical University of Munich)
CodeRetrievalGraph Neural NetworkPoint Cloud
π― What it does: This paper proposes an unsupervised 3D shape retrieval and deformation framework called U-RED, which can retrieve and deform CAD models from noisy and partially observed point clouds to closely match the target.
π― What it does: An uncertainty adaptive method for video text retrieval (UATVR) is proposed to address the multi-granularity uncertainty and one-to-many relationship issues in text-video matching.
UMC: A Unified Bandwidth-efficient and Multi-resolution based Collaborative Perception Framework
Tianhang Wang (Tongji University), Changjun Jiang (Tongji University)
CodeAutonomous DrivingOptimizationRecurrent Neural NetworkGraph Neural NetworkTime Series
π― What it does: This paper proposes a unified multi-agent collaborative perception framework (UMC) that simultaneously optimizes communication, collaboration, and reconstruction processes.
π― What it does: This paper addresses the issues of data uncertainty and model uncertainty in image tampering detection, proposing a framework based on uncertainty-guided learning, which includes an Uncertainty Estimation Network (UEN), Dynamic Uncertainty Supervision (DUS), and Uncertainty-Guided Prediction Refinement (UPR), achieving precise localization of tampered areas.
π― What it does: We propose Uni-3D, a unified model for complete 3D scene parsing and reconstruction from a single RGB image, capable of simultaneously outputting 3D semantic segmentation and geometric reconstruction of object instances and scene layout.
π― What it does: A unified coarse-fine alignment model UCOFIA is proposed for video-text retrieval, accommodating multi-level alignments of video-sentence, frame-sentence, and patch-word.
Unified Pre-Training with Pseudo Texts for Text-To-Image Person Re-Identification
Zhiyin Shao (South China University of Technology), Jingdong Wang (Baidu VIS)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageText
π― What it does: A unified pre-training pipeline called UniPT is proposed, which utilizes automatically generated pseudo-text to construct a large-scale image-text dataset LUPerson-T, addressing the issue of data and training inconsistency in text-to-image person retrieval.
Unified Visual Relationship Detection with Vision and Language Models
Long Zhao (Google Research), Ting Liu (Google Research)
CodeObject DetectionTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: Design and train a unified visual relationship detector (UniVRD) that can simultaneously perform object detection and relationship prediction in a joint label space across different datasets.
π― What it does: By inserting local temporal MHRA and global cross-attention units into the pre-trained image ViT, ViT is transformed into an efficient video understanding model called UniFormerV2.
π― What it does: This paper proposes UniSeg, a unified multi-modal LiDAR segmentation network that integrates voxel, range, and point views of RGB images and point clouds, achieving both semantic segmentation and panoptic segmentation.
π― What it does: A unified video-language temporal localization framework, UniVTG, is proposed, defining unified temporal labels (foreground, boundary, saliency) and developing a multimodal decoder that utilizes pseudo-labels for large-scale pre-training, supporting multiple tasks (moment retrieval, highlight detection, video summarization).
Unleashing the Power of Gradient Signal-to-Noise Ratio for Zero-Shot NAS
Zihao Sun (Chinese Academy of Sciences), Yu Hu (Chinese Academy of Sciences)
CodeNeural Architecture Search
π― What it does: This paper proposes a new zero-shot neural architecture search (NAS) proxy called ΞΎ-based Gradient Signal-to-Noise Ratio (ΞΎ-GSNR) for predicting network accuracy at initialization without any training.
Unleashing Vanilla Vision Transformer with Masked Image Modeling for Object Detection
Yuxin Fang (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
CodeObject DetectionTransformerImage
π― What it does: Proposes an efficient adaptation method for vanilla Vision Transformer (ViT) in object detection tasks based on Masked Image Modeling pre-training.
π― What it does: This paper proposes a method that utilizes an image foundation model (CLIP) as an unmasked teacher. By aligning unmasked visual tokens with teacher features and combining semantic masking and spatiotemporal attention, it achieves efficient unmasked video pre-training. Subsequently, cross-modal tasks are incorporated through progressive pre-training, ultimately resulting in a video foundation model capable of simultaneously handling scene, temporal, and video-text tasks.
π― What it does: A framework for 3D facial shape attribute translation based on unpaired multi-domain is proposed, capable of generating high-quality 3D facial models with different expressions, ages, and genders in one go.
π― What it does: A new unsupervised domain adaptation detection framework is proposed, which improves the generalization ability of detectors learned from labeled source domains to unlabeled target domains through network stability analysis.
π― What it does: This paper proposes an unsupervised image denoising network called SCPGabNet, which is based on self-coherent parallel generative adversarial branches and can improve denoising performance without increasing inference complexity.
π― What it does: Designed and implemented an algorithm for simultaneous data clustering and learning low-dimensional linear representations in an unsupervised mannerβManifold Linearizing and Clustering (MLC).
Unsupervised Self-Driving Attention Prediction via Uncertainty Mining and Knowledge Embedding
Pengfei Zhu (Beijing University of Posts and Telecommunications), Huadong Ma (University of Rochester)
CodeAutonomous DrivingImage
π― What it does: A novel unsupervised self-driving attention prediction model is proposed, which utilizes a natural scene pre-trained model to generate pseudo-labels and achieves adaptive training through uncertainty mining and knowledge embedding.
Jin Wang (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
CodeRestorationContrastive LearningVideo
π― What it does: This paper proposes an unsupervised video de-raining network that combines event cameras, and designs heterogeneous separation modules and cross-modal fusion modules.