Tu Le-Xuan, Trung Tran-Quang, Thi Ngoc Hien Doan, Thanh-Hai Tran
Release Date: 2022/10/19-21
3D hand pose estimation from RGB images is difficult to obtain depth information. Therefore, many studies have been conducted on 3D hand pose estimation based on 2D results. In this paper, we propose LG-Hand, a powerful method for 3D hand pose estimation that takes advantage of spatio-temporal Graph Convolutional Neural Networks (GNNs). The proposed method is characterized by incorporating spatial and temporal dependencies into a single process, and describes how dynamical knowledge plays an important role in 3D hand posture estimation. We also propose two new objective functions: angle loss, which deals with local mechanical knowledge considering the hand structure, and direction loss, which deals with global mechanical knowledge. The proposed method, LG-Hand, shows promising results on the First-Person Hand Action Benchmark (FPHAB) dataset. Various aspects of the proposed two objective functions are also discussed to demonstrate their effectiveness.
