Multiscale Attention-Based Hand Keypoint Detection

This article deals with the measurement of the hand keypoints in a vision-based setup under different constraints. Hand keypoint detection (HKD) plays a crucial role in many gesture-based applications. However, developing a generalized detection method has remained a long-standing problem. Several factors impede accurate detection: the fingers’ distance from the camera and their nearness, self-occlusion, variations in illumination, and background clutter. To overcome these barriers, we propose a two-stage architecture. The first stage generates precise hand regions, eliminating adjoining skin regions and background clutter. The second stage incorporates a novel multiscale attention block to detect keypoint coordinates precisely. Qualitative and quantitative evaluations found that the proposed architecture outperforms state-of-the-art models, with endpoint errors as low as 2.3, 1.14, and 2.11 pixels for the three benchmark datasets. This advancement lays the groundwork for future 3-D hand pose estimation developments and their applications.