Category-Aware Keypoint Masking to Address Biases in Semi-supervised 2D Pose Estimation

Xiangrui Liu, Shushi hong, Yucheng Fang, Ruirui Li

Abstract: Semi-supervised 2D human pose estimation aims to overcome the limited availability of labeled samples by leveraging a large amount of unlabeled data. However, the imbalance present in pose data often leads to data and confirmation biases, resulting in inaccurate keypoint predictions for uncommon poses by semi-supervised models. Existing research lacks a comprehensive analysis of pose data across various categories, hindering effective bias mitigation. In this study, we aim to address this obstacle by proposing an unsupervised pose clustering approach, combined with a category-aware keypoint masking technique based on difficulty levels. Our approach is straightforward, user-friendly, and can be easily integrated with other semi-supervised 2D human pose estimation methods. It notably improves the model’s performance in estimating uncommon poses.