An efficient baseline for multi-view 3d human pose estimation

Abstract

Recent advancements have been made in calculating 3D human pose keypoints from 2D joint locations obtained via a 2D backbone. While these methods demonstrate excellent performance, they demand substantial computing resources. In this work, we propose a baseline method for multi-view 3D human pose estimation using a fully connected neural network to predict 3D keypoint positions. Our approach provides a straightforward framework for fusing 2D poses from multiple camera views and regressing 3D human pose. Extensive experiments demonstrate the effectiveness of our proposed method on Human3.6M, the largest publicly available benchmark for 3D human pose estimation. Furthermore, it is important to note that increasing the number of input camera views does not inherently guarantee improved 3D pose reconstruction accuracy and quality. The optimal number of views and strategic selection of viewpoint combinations are critical factors in achieving precise 3D pose estimation results.