Abstract
3D hand pose estimation from a single depth map is an essential topic in computer vision. Most existing methods are devoted to designing a model to capture more spatial information or designing loss functions based on prior knowledge to constrain the estimated pose with prior spatial information. In this work, we focus on constraining the estimation process with spatial information adaptively by learning the mutual position relationship between joint pairs. Specifically, we propose a dynamic relationship network (DRN) with dynamic anchors. The preset fixed anchors are employed to estimate the position of each joint initially. Then, each joint is considered a dynamic anchor, which plays the role of a dynamic regressor to adjust the initially estimated position of each joint. The final estimation of each joint is the weighted sum of the results from all the dynamic anchors. Extensive experiments on benchmarks demonstrate that our method provides competitive results compared with state-of-the-arts.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have