Temporal Hints in 3D Hand Pose Estimation

Due to severe self-occlusions and complex articulated pose in hand motion, 3D hand pose estimation remains a challenging task. Most of the state-of-the-art approaches focus on how to extract spatial information effectively around the hand while ignoring the temporal consistency of hand motion, which may give hints on the locations of the occluded joints in the current frame. Inspired by solutions in action recognition, we propose to use temporal information to alleviate hand pose estimation. Specifically, we use a temporal branch to efficiently extract the temporal information of continuous image frames with negligible computational costs through convolution. By combining spatial information, our network can output more accurate hand pose estimation results. We conduct experiments on datasets with consecutive frames. Compared with methods that only use spatial information, our network achieves lower prediction errors. The experimental results also confirm our idea that the temporal information of hand motion is helpful for hand pose estimation.