Least-Squares Estimation of Keypoint Coordinate for Human Pose Estimation

Linhua Xiang, Jia Li, Zengfu Wang

Abstract

The research on human pose estimation has recently been promoted to a new high degree. As a result, existing methods that are widely used but flawed in theory must be rethought. Most researchers focus on enhancing network structure and data processing details, yet neglect to study encoding-decoding methods for keypoint coordinate. In this paper, we rethink recent encoding-decoding methods and further propose a new, elegant and reliable one. Our method is referred to as Least-squares Estimation of Keypoint Coordinate (LSEC), which is a plug-in and can be conveniently used in recent state-of-the-art (SOTA) human pose estimation models. LSEC is mathematically rigorous and unbiased, and it can compensate for the inherent bias introduced by the existing encoding-decoding methods. Besides, LSEC greatly improves the robustness of Gaussian heatmap based human pose estimation methods against adversarial attack by noise. Experiments demonstrate the effective performance and robustness of our proposed method. We will release the source code later.
BibTeX
@inproceedings{xiang2022least,
  author = {Linhua Xiang and Jia Li and Zengfu Wang},
  title = {{Least-Squares Estimation of Keypoint Coordinate for Human Pose Estimation}},
  booktitle = {PRCV},
  year = {2022},
  doi = {10.1007/978-3-031-18913-5\_35},
}

Loading related content...

Figure