Multiscale Feature Hand Pose Estimation Guided by Variable Hand Skeleton

In the realm of computer image processing, 3D hand pose estimation has found applications across various domains of human-computer interaction and virtual reality (VR). Most existing models are computationally expensive due to their spatial redundancy. In this paper, we introduce the Variable Skeleton Hand Network, a precision-focused and computationally efficient model for 3D hand joint coordinate regression. This model leverages a normalized point cloud and a 2D variable skeleton as inputs. We instantiate the model using a deformable hand skeleton in conjunction with a two-stage folding decoder. We propose the use of a multiscale single finger local decoder composed of global features and finger localized features for higher accuracy hand pose estimation. Experimental results on the ICVL and MSRA datasets show that our model achieves an error of 6.03 mm and 7.18 mm, respectively, compared to existing methods.