Unsupervised multi-view multi-person 3D pose estimation

Abstract

The problem of 3D pose estimation of multiple persons in a multi-view scenario has been an ongoing challenge in computer vision. Most current state-of-the-art methods for 3D pose estimation have relied on supervised techniques, which require a large amount of labelled data for training. However, generating accurate 3D annotations is costly, time-consuming, and prone to errors. Therefore, a novel approach that does not require labeled data for 3D pose estimation has been proposed. The proposed method, the Unsupervised Multi-View Multi- Person approach, uses a plane sweep method to generate 3D pose estimations. This approach defines one view as the target and the rest as reference views. First, the depth of each 2D skeleton in the target view is estimated to obtain the 3D poses. Then, instead of comparing the 3D poses with ground truth poses, the calculated 3D poses are projected onto the reference views. The 2D projections are then compared with the 2D poses obtained using an off-the- shelf method. Finally, the 2D poses of the same pedestrian obtained from the target and reference views are matched for comparison. The matching process is based on ground points to identify the corresponding 2D poses and compare them with the respective projections. To improve the accuracy of the proposed approach, a new reprojection loss based on the smooth L1 norm has been introduced. This loss function considers the errors in the estimated 3D poses and the projections onto the reference views. It has been tested on the publicly available Campus dataset to evaluate the effectiveness of the proposed approach. The results show that the proposed approach achieves better accuracy than state-of-the-art unsupervised methods, with a 0.5% points improvement over the best geometric system. Furthermore, the proposed method outperforms some state-of-the-art supervised methods and achieves comparable results with the best-managed approach, with only a 0.2% points difference. In conclusion, the Unsupervised Multi-View Multi-Person approach is a promising method for 3D pose estimation in multi-view scenarios. Its ability to generate accurate 3D pose estimations without relying on labeled data makes it valuable to computer vision. The evaluation results demonstrate the proposed approach’s effectiveness and potential for future research in this area.