Monocular 3D hand mesh reconstruction based on graph

3D hand pose estimation and shape reconstruction have gained considerable attention owing to advancements in deep learning techniques. This task involves estimating hand pose and mesh from images containing hands, with diverse applications across fields such as Augmented Reality (AR) and Virtual Reality (VR). However, accurately estimating and reconstructing complex hand shapes is challenging due to limitations in reconstruction models. To address this challenge, we propose a method that utilizes Graph Convolutional Networks (GCN) to progressively reconstruct 3Dhand meshes, thus improving the algorithm's accuracy and flexibility. This method takes RGB image sequences as input, leveraging the temporal relationships between adjacent frames to estimate 3D pose more accurately from2Dposes. Then, we obtain shallow sparse hand meshes from poses and employ a coarse-to-fine regression strategy to directly regress hand mesh vertices step by step. This approach not only corrects estimation errors in a timely manner, enhancing prediction accuracy, but also reduces the number of vertices regressed simultaneously, lowering hardware computational resource consumption. Our proposed method achieves competitive performance on public hand datasets FreiHAND and HO3D.