Ar3dHands: A Dataset and Baseline for Real-Time 3D Hand Pose Estimation from Binocular Distorted Images.

Hand pose estimation is an important technology for realtime human-computer interaction. Most existing methods neglect the distorted images captured by wide-angle cameras and tend to have high inference latency particularly without the acceleration of Graphic Process Units (GPUs). In this paper, we propose the first large multi-view distorted hand dataset, Ar3dHands, and develop a simple but effective 3D hand pose estimation algorithm for real-time binocular distorted images which make our method compatible with the wide-angled camera system equipped in miniature visual device like AR/VR glasses. Evaluation shows that our method can achieve state-of-the-art results on several datasets with lower mean 2D end point error and can realize real-time performance on embedded devices without GPUs.