GECM: graph embedded convolution model for hand mesh reconstruction

Hand mesh reconstruction from a single RGB image is one of the popular research topic in human understanding field with applications such as virtual/augmented reality and robot operating system. To reconstruct a hand mesh with good quality, we propose a new mesh vertex feature aggregation network module GEC. The current vertex’ features are generated by aggregating the features of the adjacent vertices according to the topological connections of the mesh vertices. Different from the traditional graph convolution structure, the GEC module circumvents the feature vectorization operation, but constructing the topological nodes with the full convolution operation. It has the advantages of avoiding destroying the spatial structure of feature maps and reducing the interference of features in the pseudo-neighborhood. Taking the GEC module as the core module, a new hand mesh reconstruction model GECM is presented. The FreiHAND dataset and the HO-3D dataset are used to evaluate the performance of the proposed GECM model. The experimental results indicate that the GECM model is superior to or on par with the state-of-the-art methods.