Domain Adaptive Hand Pose Estimation Based on Self-Looping Adversarial Training Strategy

Abstract

In recent years, with the development of deep learning methods, hand pose estimation based on monocular RGB images has made great progress. However, insufficient labeled training datasets remain an important bottleneck for hand pose estimation. Because synthetic datasets can acquire a large number of images with precise annotations, existing methods address this problem by using data from easily accessible synthetic datasets. Domain adaptation is a method for transferring knowledge from a labeled source domain to an unlabeled target domain. However, many domain adaptation methods fail to achieve good results in realistic datasets due to the domain gap. In this paper, we design a self-looping adversarial training strategy to reduce the domain gap between synthetic and realistic domains. Specifically, we use a multi-branch structure. Then, a new adversarial training strategy we designed for the regression task is introduced to reduce the size of the output space. As such, our model can reduce the domain gap and thus improve the prediction performance of the model. The experiments using H3D and STB datasets show that our method significantly outperforms state-of-the-art domain adaptive methods.