Single image based 3D human pose estimation via uncertainty learning

作者：

Highlights：

• We propose an uncertainty-based framework for 3D human pose estimation, which predicts 3D joint coordinates and uncertainties simultaneously.

• We develop an uncertainty-aware scaling factor to reshape the optimization objective. This manner improves the convergence speed and accuracy.

• We introduce a UAGCN to refine the initially estimated locations according to the estimated uncertainties.

• Experiments show the effectiveness of our approach quantitatively and qualitatively. In noisy scenes, our model exhibits greater robustness.

摘要

•We propose an uncertainty-based framework for 3D human pose estimation, which predicts 3D joint coordinates and uncertainties simultaneously.•We develop an uncertainty-aware scaling factor to reshape the optimization objective. This manner improves the convergence speed and accuracy.•We introduce a UAGCN to refine the initially estimated locations according to the estimated uncertainties.•Experiments show the effectiveness of our approach quantitatively and qualitatively. In noisy scenes, our model exhibits greater robustness.

论文关键词：Uncertainty,3D pose estimation,Graph convolutional network

论文评审过程：Received 10 August 2020, Revised 26 April 2022, Accepted 25 July 2022, Available online 26 July 2022, Version of Record 4 August 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2022.108934