Unsupervised Shape Enhancement and Factorization Machine Network for 3D Face Reconstruction

Existing unsupervised methods are often unable to capture accurate 3D shapes due to the ambiguity of shapes and albedo maps, limiting their applicability to downstream tasks. Therefore, this article proposes an unsupervised shape enhancement and decomposition machine network for 3D facial reconstruction. Specifically, we design a shape enhancement network, further combining global and local features, which can restore more complete and realistic albedo images without introducing additional supervision, so as to obtain higher-quality 3D faces. Secondly, based on the principle of decomposition machines, we design a decomposition module. By decomposing large matrices, the network learns to infer better results, while reducing the number of network parameters further improving the accuracy of our model. Extensive experiments on BFM and CelebA data demonstrate the effectiveness of our methods.