Information Fusion Cascade Network for 3D Hand Pose Estimation Using Depth Data

With the development of metaverse, augmented reality and robot operation, hand pose estimation plays an increasingly important role. In this paper, an information fusion cascade network is proposed to improve the performance of hand pose estimation through a cascade network containing finger-palm information fusion and a biomechanical-constraint loss function. Extensive evaluations on two public datasets show that information fusion cascade network can generate accurate and fast 3D hand poses and achieve an excellent balance of 3D hand pose estimation accuracy and speed compared to state-of-the-art (SOAT) methods.