Hourglass Network for Hand Pose Estimation with RGB Images

Recent studies have shown effectiveness in using RGB images to estimate 3D gestures. However, due to hand flexibility and occlusion problems, it is very difficult to achieve the best results. In the current research, to estimate 3D gestures from RGB images, it is necessary to locate the 2D coordinate position of each key point firstly, and then derive the 3D posture. Therefore, it is very important for the final pose estimation accurately to locate the 2D coordinate positions of gestures. In this paper, we propose to use Hourglass Network to locate the 2D coordinates of each key point. The Hourglass Network locates the key points from the whole to the local, which has a good effect on grasping the relationship between the key points. The experimental results show that the method effectively improves the accuracy of 2D coordinate estimation, and thus makes the best effect in the estimation of 3D pose, and evaluates it on the sign language identification dataset.