LightHand99K: A Synthetic Dataset for Hand Pose Estimation With Wrist-Worn Cameras

Abstract

Although hand-pose estimation using external camera systems has made significant progress driven by large annotated datasets, wrist-worn camera-based hand-pose estimation offers unique advantages owing to its ability to capture nearby images. However, datasets primarily collected from multi-angle external systems are unsuitable for wrist-camera-based estimation. To address this gap, we developed LightHand99K, a synthetic dataset comprising 99,792 synthetic RGB hand images designed explicitly for hand-pose estimation using wrist-worn cameras. The dataset was generated using a Unity 3D-based hand image generator optimized to efficiently enhance the performance of hand-pose estimation models for wrist-worn camera images by incorporating human finger joint movement information while minimizing computational cost. This tool allows users to customize their poses, camera angles, backgrounds, lighting, and hand-orientation settings. Incorporating three data augmentation techniques, LightHand99K demonstrated a 36% increase in area under the curve (AUC) and a 6.2-mm reduction in average endpoint error (EPE) compared to existing datasets. These results underscore the value of LightHand99K in advancing hand-pose estimation, particularly in wrist-worn camera applications.