Generating Hand Posture and Motion Dataset for Hand Pose Estimation in Egocentric View

Abstract

Hand interaction is one of the main input modalities for augmented reality glasses. Vision-based approaches using deep learning have been applied to hand tracking and shown good results. To train a deep neural network, a large dataset of hand information is required. However, obtaining real hand data is painful due to a large number of annotations and lack of diversities such as skins, lighting conditions, and backgrounds. In this paper, we propose a method to generate a synthetic hand dataset that includes diverse human and environmental parameters. By applying constraints of a human hand, we can get realistic hand poses for hand dataset. We also generate dynamic hand animations which can be used for hand gesture recognition.