🎯 What it does: A multi-frame point cloud interpolation method based on a 4D spatiotemporal neural field, NeuralPCI, is proposed to address the problem of 3D point cloud interpolation under nonlinear large motion.
NewsNet: A Novel Dataset for Hierarchical Temporal Segmentation
Haoqian Wu (Tencent), Bernard Ghanem (National Tsing Hua University)
CodeSegmentationTransformerVideoMultimodality
🎯 What it does: The researchers constructed a large-scale news video dataset called NewsNet and conducted a hierarchical temporal segmentation study on this dataset.
🎯 What it does: This paper proposes a 3D GAN framework that can unsupervisedly learn high-quality, 3D-consistent, and animatable facial avatars from unstructured 2D images, supporting fine-grained control over full head rotation, expressions, eye blinking, gaze, and achieving high-fidelity rendering.
🎯 What it does: This paper presents a large-scale, domain-wide NICO++ dataset and provides a more reasonable evaluation protocol for the domain generalization problem.
NIKI: Neural Inverse Kinematics With Invertible Neural Networks for 3D Human Pose and Shape Estimation
Jiefeng Li (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
CodePose EstimationFlow-based ModelImage
🎯 What it does: A method for inverse kinematics based on reversible neural networks, NIKI, is proposed for robust and pixel-aligned 3D human pose and shape estimation from monocular images.
🎯 What it does: A noise proxy-based integrated pseudo-quantization (NIPQ) method is proposed, which achieves quantization-aware training of low-precision networks by replacing STE with pseudo-quantization noise during the training process, and supports unified quantization of weights and activations as well as mixed precision scheduling.
Noisy Correspondence Learning With Meta Similarity Correction
Haochen Han (Xi'an Jiaotong University), Minnan Luo (Xi'an Jiaotong University)
CodeRetrievalMeta LearningImageText
🎯 What it does: This study investigates similarity correction and model robustness enhancement in image-text retrieval under the presence of noise through a meta-learning framework.
CodeRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper explores the introduction of non-contrastive learning methods into language-image pre-training and proposes the xCLIP multi-task framework, which integrates the advantages of CLIP and nCLIP to achieve stronger zero-shot transfer and representation learning.
🎯 What it does: This paper proposes a new category discovery framework NOPS for 3D point cloud semantic segmentation based on online clustering and uncertainty quantification, which can automatically identify and segment unlabelled new category points under the premise of having only base category labels.
🎯 What it does: A language-guided image inpainting method called N WA-LIP is proposed, which can fill in missing areas based on text descriptions while keeping the undamaged regions unchanged.
🎯 What it does: A self-supervised scene-adaptive object detection framework is proposed, utilizing a pre-trained detector to generate pseudo-labels combined with tracking, location-aware artifact-free Mixup, and dynamic background extraction, significantly improving detection accuracy in fixed camera scenarios.
🎯 What it does: Transform heatmap keypoint detection into a statistically reliable set of circle/ellipse predictions, and propagate these uncertainties to 6D poses, providing a computable worst-case error upper bound.
Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection
Luting Wang (Beihang University), Si Liu (Beihang University)
CodeObject DetectionKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelImage
🎯 What it does: Proposes the Object-Aware Distillation Pyramid (OADP) framework to extract complete and pure object knowledge from the pre-trained Vision-and-Language model (CLIP) for open vocabulary object detection, and transfers this knowledge to the Faster R-CNN detector through a three-level distillation (global, block-level, object-level).
Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking
Jinkun Cao (Carnegie Mellon University), Kris Kitani (Carnegie Mellon University)
CodeObject TrackingAutonomous DrivingVideo
🎯 What it does: This paper proposes an observation-centric multi-object tracking framework (OC-SORT) that improves robustness against occlusion and nonlinear motion through observation-driven re-update and momentum consistency based on Kalman filtering.
Mehdi Zemni (Valeo AI), Matthieu Cord (Sorbonne Université)
CodeAutonomous DrivingExplainability and InterpretabilityAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an object-oriented adversarial explanation framework called OCTET, designed to generate interpretable adversarial examples for visual models.
🎯 What it does: This paper proposes a single Vision Transformer (ViT) model called OmniMAE, which jointly pre-trains on both image and video visual modalities through unsupervised masked autoencoding (MAE) and directly transfers to downstream tasks.
🎯 What it does: For unpaired infrared-visible video translation, the CPTrans framework is proposed to achieve fine-grained content-rich patch transfer.
🎯 What it does: This paper proposes the DeepMVC framework to unify deep multi-view clustering methods and analyzes the impact of self-supervised learning (especially contrastive alignment) on clustering performance, pointing out that alignment reduces the number of distinguishable clusters as the number of views increases.
🎯 What it does: A multi-modal high-precision dataset is proposed, and based on this, the impact of noise from different depth sensors (I-ToF, D-ToF, active stereo, RGB+P, etc.) on dense 3D vision tasks (monocular depth estimation, implicit reconstruction, etc.) is evaluated.
🎯 What it does: In the open, online visual classification scenario, an Instant Category Discovery (OCD) task is proposed, utilizing a scalable recognition model with hash coding to achieve instant recognition of new and old categories.
🎯 What it does: A one-to-few (label assignment) strategy is proposed, utilizing soft anchors to dynamically adjust positive and negative weights, enabling end-to-end dense detection without NMS in fully convolutional networks.
🎯 What it does: This paper proposes OneFormer, a unified Transformer-based image segmentation framework that can simultaneously perform semantic, instance, and panoptic segmentation in a single model with a single training session.
🎯 What it does: This paper proposes a parameter-free upsampling module called OPE-Upscale, which utilizes Orthogonal Position Encoding (OPE) to achieve continuous reconstruction of images at arbitrary scales, completely replacing traditional MLP-based implicit networks.
🎯 What it does: A transfer method for few-shot open set recognition, OSLO, is proposed, which introduces a latent inlierness score within a maximum likelihood framework to simultaneously perform category prediction and anomaly detection.
Open-Vocabulary Point-Cloud Object Detection Without 3D Annotation
Yuheng Lu (Peking University), Shanghang Zhang (University of California Berkeley)
CodeObject DetectionTransformerVision Language ModelContrastive LearningPoint Cloud
🎯 What it does: Utilize a 2D pre-trained detector to generate pseudo 3D boxes and learn a 3D detector, then achieve text-point cloud cross-modal alignment through CLIP, enabling open vocabulary point cloud detection without 3D annotations.
🎯 What it does: A unified and scalable gait recognition codebase, OpenGait, has been constructed, and existing methods have been fairly reproduced and comprehensively ablated on four mainstream datasets, ultimately proposing a structurally simple and powerful baseline model, GaitBase.
Optimal Transport Minimization: Crowd Localization on Density Maps for Semi-Supervised Counting
Wei Lin (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
CodeOptimizationImage
🎯 What it does: This paper proposes a parameter-free Optimal Transport Minimization (OT-M) algorithm that converts density maps into hard point annotations and is used for semi-supervised counting;
🎯 What it does: This paper proposes an 'orthogonal labeling' method that only requires labeling two mutually perpendicular slices in each 3D medical image volume, and based on this, constructs a Dense-Sparse Co-Training (DeSCO) model to achieve efficient semi-supervised segmentation under sparse labeling.
OSRT: Omnidirectional Image Super-Resolution With Distortion-Aware Transformer
Fanghua Yu (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Chao Dong (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
CodeRestorationSuper ResolutionTransformerImage
🎯 What it does: For the super-resolution task of panoramic (360°) images, the authors propose fisheye downsampling to simulate the real imaging process and design a distortion-aware Transformer (OSRT) that adaptively modulates ERP distortion through two modules, DAAB and DACB; at the same time, an augmentation strategy is employed to generate pseudo ERP data from planar images to alleviate overfitting in large models.
🎯 What it does: This paper proposes a one-shot 3D consistent talking head system called OTAvatar, which utilizes controllable three-plane rendering to generate high-quality animated avatars from a single portrait, allowing for any viewpoint, expression, and pose.
🎯 What it does: This paper proposes a framework called OOD Semantic Pruning (OSP) for removing OOD information at the semantic level, aimed at improving ID classification and OOD detection performance in robust semi-supervised learning.
PA&DA: Jointly Sampling Path and Data for Consistent NAS
Shun Lu (Chinese Academy of Sciences), Chengru Song (Kuaishou Technology)
CodeNeural Architecture SearchImage
🎯 What it does: Proposes the PA&DA method, which jointly optimizes the path and data sampling distribution in supernet training to reduce gradient variance and improve ranking consistency in one-shot NAS.
🎯 What it does: A unified large-scale dataset called PACO has been constructed, which includes objects, object parts, and attributes, and three benchmark tasks are provided: part segmentation, object/part attribute prediction, and zero-shot instance retrieval.
Paired-Point Lifting for Enhanced Privacy-Preserving Visual Localization
Chunghwan Lee (Hanyang University), Je Hyeong Hong (Hanyang University)
CodePose EstimationSafty and PrivacySimultaneous Localization and MappingPoint Cloud
🎯 What it does: Designed and implemented the Paired-Point Lifting (PPL) scheme, which pairs sparse point clouds and connects them into line clouds, thereby enhancing privacy protection while maintaining visual localization accuracy.
🎯 What it does: This paper proposes a panoramic scene graph generation task for videos (PVSG) and constructs a corresponding dataset and benchmark model.
🎯 What it does: This paper studies the blind inverse problem and proposes the BlindDPS parallel diffusion backward sampling method to achieve joint estimation of the operator and the image.
🎯 What it does: A Partial Network Cloning method is proposed to clone transferable modules from pre-trained models and directly insert them into target models, allowing the target model to acquire new functionalities without altering the original parameters.
🎯 What it does: An unsupervised image denoising framework called Patch-Craft is proposed, which generates artificially synthesized 'cropped' target images by performing nearest neighbor matching (excluding itself) on overlapping patches of each input image within a group of images. These targets are then used as training labels to learn the denoiser; during training, statistical analysis of the target noise is conducted, and low covariance samples are trimmed to reduce the correlation between target noise and input noise.
🎯 What it does: A novel online class-incremental learning framework called PCR, which combines proxy and contrastive learning, is proposed to alleviate the problem of catastrophic forgetting.
🎯 What it does: A post-training quantization method based on prediction difference measurement, PD-Quant, is developed to optimize quantization parameters and reduce overfitting through distribution correction.
🎯 What it does: The study generates goal-oriented action sequences given the initial and target visual states in instructional videos, proposing to view process planning as conditional distribution fitting, using a projection diffusion model to output the complete action sequence in one go.
🎯 What it does: A semi-supervised medical image classification framework called PEFAT is proposed, which combines pseudo-loss estimation and feature-level adversarial training to improve classification performance with limited labeled data.
Perception and Semantic Aware Regularization for Sequential Confidence Calibration
Zhenghua Peng (South China University of Technology), Shuangping Huang (Guangdong University of Technology)
CodeRecognitionRecurrent Neural NetworkLarge Language ModelTextSequentialAudio
🎯 What it does: This paper proposes a Perceptual and Semantic Perception Sequence Regularization framework (PSSR), which enhances the confidence calibration of deep sequence recognition models by introducing sequences that are visually similar and semantically related to the target sequence as additional supervision.
🎯 What it does: A single image super-resolution framework is proposed, which utilizes Optimal Objective Estimation (OOE) to dynamically select the optimal loss combination for each region, and achieves variable target super-resolution through a generative model.
🎯 What it does: A pose-guided portrait generation framework PIDM based on a denoising diffusion model is proposed, which can synthesize high-quality portraits under the conditions of a given pose and source image.
🎯 What it does: In the pedestrian re-identification task, the authors propose Patch-wise High-frequency Augmentation (PHA), which enhances the Vision Transformer's ability to express high-frequency details and improves its recognition performance through frequency domain decomposition and self-attention analysis.
Phase-Shifting Coder: Predicting Accurate Orientation in Oriented Object Detection
Yi Yu (Southeast University), Feipeng Da (Southeast University)
CodeObject DetectionImage
🎯 What it does: A phase shift encoder (PSC) and its dual-frequency version (PSCD) are proposed for angle regression in inclined target detection, addressing boundary discontinuity and square issues;
Physics-Guided ISO-Dependent Sensor Noise Modeling for Extreme Low-Light Photography
Yue Cao (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
CodeRestorationFlow-based ModelImagePhysics Related
🎯 What it does: This study investigates the formation mechanism of camera noise in extremely low light environments, proposing a physics-guided ISO-dependent noise model implemented within a normalizing flow framework.
Pic2Word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval
Kuniaki Saito (Boston University), Tomas Pfister (Google Cloud AI Research)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: The Pic2Word model is proposed, which utilizes weakly labeled image-caption pairs and unlabeled images to train under the CLIP framework, mapping images to language word vectors to achieve zero-shot synthesized image retrieval.
🎯 What it does: A three-branch network called PIDNet has been developed, which integrates detail, context, and boundary information to achieve real-time semantic segmentation.
🎯 What it does: This paper re-evaluates local aggregators in LiDAR point cloud 3D object detection and proposes the PillarNeXt network based on pillar encoders, combining 2D detection network architecture and training strategies, ultimately achieving state-of-the-art performance on Waymo and nuScenes.
🎯 What it does: A cross-modal pre-training framework PiMAE based on masked autoencoders is proposed, which jointly learns representations of point clouds and RGB images and fine-tunes them under multiple tasks.
PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification
Meike Nauta (University of Twente), Christin Seifert (University of Duisburg-Essen)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: An interpretable image classification network PIP-Net is proposed, which utilizes interpretable prototype segments to achieve class decisions.
🎯 What it does: This paper proposes a planning-oriented integrated autonomous driving framework called UniAD, which integrates perception, prediction, and planning within a unified Transformer system, using a query interface to connect various modules and achieve end-to-end learning.
🎯 What it does: This paper proposes a paired mask image modeling (pMIM) pre-training method for the encoder and decoder of dense geometric matching networks, significantly enhancing cross-view feature alignment capabilities.
🎯 What it does: A prototype-based modality rebalancing (PMR) strategy is proposed, which accelerates slow-learning modalities using a non-parametric prototype classifier and alleviates the suppression of dominant modalities through prototype entropy regularization, thereby addressing the modality imbalance issue in multimodal learning.
🎯 What it does: A generative model based on Polynomial Implicit Neural Representation (Poly-INR) is proposed, capable of generating high-resolution images on large and diverse datasets.
🎯 What it does: A pose-disentangled contrastive learning framework (PCL) is proposed, which achieves the separation of pose-related and pose-agnostic facial features through a Pose-Disentangled Decoder and is trained in a self-supervised facial representation.
PoseExaminer: Automated Testing of Out-of-Distribution Robustness in Human Pose and Shape Estimation
Qihao Liu (Johns Hopkins University), Alan L. Yuille (Max Planck Institute for Informatics)
CodePose EstimationReinforcement LearningImage
🎯 What it does: An automated testing framework named PoseExaminer has been developed for the systematic evaluation of the robustness of human pose and shape estimation methods in out-of-distribution (OOD) scenarios.
Position-Guided Text Prompt for Vision-Language Pre-Training
Jinpeng Wang (National University of Singapore), Shuicheng Yan (Sea AI Lab)
CodeObject DetectionRetrievalTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a Position-guided Text Prompt (PTP) to enhance the visual localization and alignment capabilities of Visual-Language Pre-training (VLP) models. The method divides images into N×N blocks, generates object labels for each block using an object detector or the CLIP model, and transforms the visual localization task into a fill-in-the-blank problem by inserting position placeholders in the text (e.g., "The block [P] has a [O]"). This allows the model to learn finer-grained spatial information during the pre-training phase.
🎯 What it does: A model-free, training-free post-processing method called Gaussian Approximated Post-processing (GAP) is proposed to correct time quantization errors caused by preprocessing in temporal action detection.
🎯 What it does: A content-aware visual text layout generation method is proposed, which can automatically arrange elements such as text, logos, and backgrounds on an existing content canvas.
Power Bundle Adjustment for Large-Scale 3D Reconstruction
Simon Weber (Technical University of Munich), Daniel Cremers (University of Oxford)
CodeOptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A Bundle Adjustment solver PoBA based on the inverse Schur complement power series expansion is proposed, achieving fast and low-memory solutions for large-scale 3D reconstruction problems.
🎯 What it does: A framework for image registration called PRISE is proposed, which optimizes the deep LK method through strong star convex constraints.
Privacy-Preserving Representations Are Not Enough: Recovering Scene Content From Camera Poses
Kunal Chelani (Chalmers University of Technology), Zuzana Kukelova (Czech Technical University in Prague)
CodeObject DetectionPose EstimationSafty and PrivacySimultaneous Localization and MappingImage
🎯 What it does: By sending a large number of single object images to a cloud-based visual localization service, the returned camera pose information is used to infer the type and location of objects in the scene, thereby demonstrating that even with privacy-preserving representations, pose alone can leak scene content.
🎯 What it does: A probabilistic model-based objectness estimation framework (PROB) is proposed and integrated into Deformable DETR for open-world object detection.
Probabilistic Knowledge Distillation of Face Ensembles
Jianqing Xu (Tencent), Bryan Hooi (National University of Singapore)
CodeRecognitionKnowledge DistillationImage
🎯 What it does: This paper proposes the Bayesian Ensemble Averaging (BEA) method, which extends traditional mean ensemble through probabilistic modeling, and based on this, designs the BEA-KD knowledge distillation framework that compresses the probabilistic embeddings from multiple teachers into a single student network.
🎯 What it does: A probability-based global cross-modal upsampling method (PGCU) for pansharpening is proposed, which fully utilizes the global information of low-resolution multispectral images (LRMS) and the cross-modal information of panchromatic images (PAN), while considering channel specificity.
🎯 What it does: This paper proposes an emotion-oriented pre-training method based on the human visual emotional perception mechanism, dividing the pre-training process into three stages (stimulus acquisition, overall organization, and advanced perception), and training independent models at each stage; subsequently, the emotional knowledge from multiple models is integrated into a single target model through feature and logits regularization to enhance visual emotion analysis (VSA) performance.
Progressive Neighbor Consistency Mining for Correspondence Pruning
Xin Liu (Nankai University), Jufeng Yang (Nankai University)
CodePose EstimationGraph Neural NetworkImage
🎯 What it does: A Neighborhood Consistency Mining Network (NCMNet) is proposed for correspondence pruning, which can efficiently identify inliers in feature matching and estimate camera pose.
🎯 What it does: A framework for Open Set Model Attribution (OSMA) is proposed, utilizing a progressively expanded lightweight gain model to simulate the fingerprints of unknown models within a known model space, achieving simultaneous identification of both known and unknown generative models.
🎯 What it does: In the task of generalized zero-shot learning, this paper proposes a Progressive Semantic-Visual Mutual Adaption (PSVMA) network, which utilizes a Dual Semantic-Visual Transformer Module (DSVTM) to progressively learn instance-centered attribute prototypes and generate unambiguous visual representations to enhance recognition of unseen categories.
🎯 What it does: A progressive spatiotemporal alignment framework based on event cameras is proposed, which can efficiently estimate motion parameters for rotation, homography, and 6-DOF motion models.
Prompt, Generate, Then Cache: Cascade of Foundation Models Makes Strong Few-Shot Learners
Renrui Zhang (Shenzhen Institute of Advanced Technology Chinese Academy of Science), Hongsheng Li (Chinese University of Hong Kong)
CodeClassificationGenerationTransformerLarge Language ModelPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes CaFo Cascaded Foundation Models, which integrate the pre-trained knowledge of CLIP, DINO, DALL-E, and GPT-3 through a three-step process of Prompt-Generate-Cache to achieve few-shot image classification.
Prompting Large Language Models With Answer Heuristics for Knowledge-Based Visual Question Answering
Zhenwei Shao (Hangzhou Dianzi University), Jun Yu (Hangzhou Dianzi University)
CodeTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: By generating answer candidates and answer-related examples as two types of heuristics and embedding them into the prompt, we utilize GPT-3 to perform reasoning on knowledge-based VQA tasks.
🎯 What it does: This paper proposes a Proposal-based Multiple Instance Learning (P-MIL) framework for weakly supervised temporal action localization, which directly classifies candidate proposals to eliminate the inconsistency between training and testing objectives in traditional S-MIL.
🎯 What it does: This paper proposes the Prototype-based Embedding Network (PE-Net), which maps entities and predicates into a semantic space based on word vectors and performs gated fusion of instance features to generate compact and distinguishable representations, achieving entity-predicate matching to enhance the relationship recognition performance in Scene Graph Generation (SGG).
🎯 What it does: A white-box adversarial attack method for semantic segmentation models, ALMA prox, is proposed, which can generate extremely small ∞-norm perturbations.
🎯 What it does: This paper proposes a single-stage 3D object detection framework PVT-SSD, which integrates voxel and point cloud representations and utilizes a point-voxel Transformer for efficient detection.
🎯 What it does: This paper proposes a low-bit quantization method for DETR (Q-DETR) and corrects the information distortion of the quantized query vectors through Distributional Correction Distillation (DRD), significantly improving the detection performance of low-bit DETR.
CodeDomain AdaptationData-Centric LearningTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: A teacher model is constructed using a large-scale visual language model (VLM), leveraging unlabeled image generation question-answer pseudo-labels to enhance self-training on the target small-scale VQA dataset;
🎯 What it does: This paper studies a motion matching framework based on quantization and phase guidance for generating natural and coherent gestures from speech and text.
Matteo Farina (University of Trento), Federica Arrigoni (Politecnico di Milano)
CodeOptimizationTabularPhysics Related
🎯 What it does: The first quantum method QUMF for multi-model geometric fitting using decoherent quantum computers is proposed, along with a decomposable version DEQUMF to address large-scale problems.
🎯 What it does: This paper proposes a method that adds small Gaussian noise to model weights and utilizes Taylor expansion to achieve randomized adversarial training, aiming to enhance both the model's adversarial robustness and clean data accuracy.
🎯 What it does: This paper studies the vulnerability of Spiking Neural Networks (SNN) to adversarial attacks and proposes a new attack method called Rate Gradient Approximation Attack (RGA).
Fidel A. Guerrero Peña (ETS Montreal), Marco Pedersoli (ETS Montreal)
CodeOptimizationImage
🎯 What it does: This paper proposes a differentiable re-basin method based on the Sinkhorn operator and implicit differentiation to find function-equivalent rearrangements in the parameter space of neural networks, enabling model transfer, fusion, and incremental learning.
🎯 What it does: This paper proposes Re-GAN, a GAN training framework that dynamically prunes and regrows during the training process, enabling the exploration of different subnetwork structures under limited data, thus achieving more efficient training.
Re-Thinking Federated Active Learning Based on Inter-Class Diversity
SangMook Kim (KAIST AI), Se-Young Yun (KAIST AI)
CodeFederated LearningImage
🎯 What it does: The LoGo algorithm is proposed within the framework of federated active learning, combining global and local models for a two-step clustering query, aiming to address both local heterogeneity and global imbalance issues.
🎯 What it does: This paper proposes a reversible network reconnection method (Re 2 TAL) that transforms existing pre-trained video backbone networks into reversible networks, achieving end-to-end temporal action localization, significantly reducing memory usage and improving performance.
🎯 What it does: The first unedited, field video dataset for multi-person eye blink detection, MPEblink, is proposed, and the InstBlink model is introduced based on a single-stage query Transformer, achieving simultaneous face detection, tracking, and instance-level blink detection.
REC-MV: REconstructing 3D Dynamic Cloth From Monocular Videos
Lingteng Qiu (Chinese University of Hong Kong, Shenzhen), Xiaoguang Han (Chinese University of Hong Kong, Shenzhen)
CodeGenerationOptimizationVideoMesh
🎯 What it does: By jointly optimizing explicit feature curves and implicit SDF from monocular video, we extract temporally coherent dynamic 3D garment meshes with open boundaries.