Enhancing 3D hand pose estimation using SHaF: synthetic hand dataset including a forearm

Currently, there is an increased need for training images in 3D hand pose estimation and a higher reliance on computationally intensive 3D mesh annotations for 3D coordinate estimations. Considering this, this study introduces a new hand image dataset called Synthetic Hand Dataset Including a Forearm (SHaF) and an efficient transformer-based three-dimensional (3D) hand pose estimation model tailored to extract hand postures from hand images. The proposed dataset comprises diverse synthetic hand posture images, across various cameras and environmental settings, which were generated using the Unity 3D hand model. It differs from existing artificial hand datasets in that it includes the forearm in its synthetic images. Given that real-world hand images often capture both the hand and forearm, our dataset bolsters the accuracy of hand pose estimation in practical scenarios. Regarding the proposed model, it uses the pose graph module (PGM) and auxiliary pose estimation module (APEM), thereby offering efficient 3D hand pose estimation without requiring 3D mesh information. Through comparative experiments with established datasets and models in hand pose estimation as well as various ablation studies, we confirmed the efficacy of our dataset and the superior performance of the estimation model over that of other methods.