Faster, Better and More Detailed: 3D Face Reconstruction with Graph Convolutional Networks

This paper addresses the problem of 3D face reconstruction from a single image. While available solutions for addressing this problem do exist, to our knowledge, we propose the very first approach which is robust, lightweight and detailed i.e. it can reconstruct fine facial details. Our method is extremely simple and consists of 3 key components: (a) a lightweight non-parametric decoder based on Graph Convolutional Networks (GCNs) trained in a supervised manner to reconstruct coarse facial geometry from image-based ResNet features. (b) An extremely lightweight (35K parameters) subnetwork - also based on GCNs - which is trained in an unsupervised manner to refine the output of the first network. (c) A novel feature-sampling mechanism and adaptation layer which injects fine details from the ResNet features of the first network into the second one. Overall, our method is the first one (to our knowledge) to reconstruct detailed facial geometry relying solely on GCNs. We exhaustively compare our method with 7 state-of-the-art methods on 3 datasets reporting state-of-the-art results for all of our experiments, both qualitatively and quantitatively, with our approach being, at the same time, significantly faster.