Monocular 3D Human Pose Estimation via Euler Angles

Jianchu Lin, Du Guangxiao, FuSong, Xiaobing Chen, Shangbing Gao, Yan Yunyang, Yongtao Yu, Xin Ning

Keywords: 3D human pose estimation, Joint rotation, Euler angle, body's self-rotation discontinuities, orientation conditions

TL;DR: 3D human pose estimation using Euler angles addresses the problem of body's discontinuous rotations by pre-estimated conditions.

Abstract: Monocular 3D human pose estimation is a key challenge in computer vision. While existing joint position-based methods often struggle with accurate bone length prediction and rotation ambiguities under joint collinearity, joint rotationbased methods circumvent the bone length issue but face discontinuity problems when regressing the body’s self-rotation angle, limiting their practical application. In this work, we theoretically analyze the root cause of this discontinuity and propose a framework for Euler angle-based pose estimation. Our method transforms Euler angle into its corresponding sine-cosine representation on the unit circle, ensuring a continuous and wrap-around-free formulation, and introduces body orientation as an internally learned conditional variable, which jointly optimizes the prediction of Euler angles in a unified learning process. By allowing the network to adapt its regression behavior based on the predicted condition, the framework effectively resolves the discontinuity inherent in Euler-angle regression. Experiments across diverse model architectures, including CNNs, GCNs, and Transformers, demonstrate that our approach enables continuous and stable self-rotation prediction. The framework forms a versatile and efficient module that maintains compatibility with existing backbones and facilitates streamlined end-to-end training.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17390