Read the original document by opening this link in a new tab.
Table of Contents
1 Introduction
2 Related Work
2.1 2D Keypoint Localization Dataset
2.2 Keypoints Localization Method
3 COCO-WholeBody Dataset
3.1 Data Annotation
3.2 Evaluation Protocol and Evaluation Metrics
Summary
This paper investigates the task of 2D human whole-body pose estimation, which aims to localize dense landmarks on the entire human body including face, hands, body, and feet. The authors introduce COCO-WholeBody, which extends the COCO dataset with whole-body annotations, providing a benchmark with manual annotations on the entire human body. They propose a single-network model, ZoomNet, to solve the scale variation of different body parts and demonstrate its significant performance on the COCO-WholeBody dataset. The dataset enables training deep models for whole-body pose estimation and serves as a pre-training dataset for tasks like facial landmark detection and hand keypoint estimation. The authors highlight the importance of whole-body pose estimation for applications like virtual reality and action recognition. They discuss dataset biases and the challenges of existing datasets in comprehensive human pose estimation. The proposed ZoomNet outperforms previous methods by effectively handling the scale variance in whole-body pose estimation. The authors present detailed annotation protocols, evaluation metrics, and quality control measures for the COCO-WholeBody dataset.