End-to-End 3D Facial Shape Reconstruction From an Unconstrained Image
Yanghong Han, Xing Qiao, Yan Xiang Wu, Zili Zhang · 2021
Three-dimensional face generation is of high importance for computer graphics and computer vision applications. The reconstruction task of human faces is challenging and face varies extensively when considering pose, expression, occlusion and illumination. Conventional face models are learned from a set of 3D face scans and represented by Principal Component Analysis(PCA), which is always constrained in linear space. Nevertheless, the shape of face and variations in images cannot be represented faithfully by these traditional models, which are very crucial for face synthesis. To address this problem, a novel architecture is proposed to learn face models in nonlinear spaces. Explicitly, our framework is a Convolutional Neural Networks(CNN) based End-to-End 3D Facial Shape Reconstruction(3DFSR-E2E) from a single-view 2D image scheme. And the variations in facial images are also coped with the model. Firstly, our Encoder is composed of new Improved Residual Blocks(IR-B) utilized in face recognition task, and Decoder consists of Fractionally-Strided Convolutions. Secondly, a UV map is estimated by the trainable Encoder-Decoder network, which is a representation of face geometry and face alignment information. Furthermore, a novel loss function is combined by a weighted Vertex-wise and a Laplacian regularization loss. The measure is employed to improve the results of reconstructed face model. Finally, the experimental results demonstrate the benefits and effectiveness of our overall method with qualitative and quantitative comparisons on public challenging face datasets.