Normalized Triangulation for Calibrated Dual-View 3D Human Pose Estimation

Abstract

In this work, we decouple calibrated dual-view 3D human pose estimation (HPE) into the well-studied problems of 2D pose estimation, and 2D-to-3D pose lifting, focusing on the latter task. The key challenges stem from: 1) 2D pose is noisy and unreliable due to occlusion and motion blur, and 2) the trained model cannot generalize well to unseen camera configurations. To overcome these limitations, we propose three interconnected innovations: First, a Normalized Triangulation that transforms the 2D pose from pixel space to 3D normalized rays, which makes our approach robust to the camera parameters change. Second, a hybrid neural-geometry framework (i.e., including refinement and triangulation) that explicitly incorporates multi-view geometry into our models. Third, an analytical inverse kinematics (AnalyIK) solver that decomposes articulated motion with human topology, which simultaneously considers symmetry constraint and joint angle limit. Experiments show that the proposed framework achieves state-of-the-art performance on two widely used benchmarks (i.e., Huamn3.6M and HumanEva-I). Code is available at: https://github.com/Z-Z-J/Normalized-Triangulation.