3D Human Pose Estimation with Siamese Equivariant Embedding

TL;DR

This paper introduces a siamese neural network architecture that learns rotation-equivariant embeddings to improve monocular 3D human pose estimation, reducing overfitting to camera angles and achieving state-of-the-art cross-camera accuracy.

Contribution

The authors propose a novel siamese network with rotation-equivariant embeddings that enhances 3D pose estimation robustness across different camera views.

Findings

Consistent error reduction across multiple datasets

State-of-the-art cross-camera error rate

Effective with various base networks

Abstract

In monocular 3D human pose estimation a common setup is to first detect 2D positions and then lift the detection into 3D coordinates. Many algorithms suffer from overfitting to camera positions in the training set. We propose a siamese architecture that learns a rotation equivariant hidden representation to reduce the need for data augmentation. Our method is evaluated on multiple databases with different base networks and shows a consistent improvement of error metrics. It achieves state-of-the-art cross-camera error rate among algorithms that use estimated 2D joint coordinates only.