Face manifold: manifold learning for synthetic face generation

The face is a crucial aspect of human communication and identity. Accurately estimating face structure is a fundamental task in computer vision, with significant applications in various fields, including facial recognition and medical surgeries. Deep learning techniques have made notable progress in 3D face reconstruction from 2D images. However, this approach demands large 3D face datasets, often tackled by synthetic face generation. Unfortunately, synthetic datasets can contain non-possible faces, which pose significant challenges. This paper presents a novel approach to synthetic diverse face dataset generation by leveraging face manifold learning. We divide the face structure into shape and expression groups and use a fully convolutional autoencoder network to handle non-possible faces while preserving dataset diversity. The proposed method is used to train deep 3D reconstruction networks and results indicate that our proposed method demonstrates its effectiveness in denoising highly corrupted faces. Additionally, we assess the diversity of the generated dataset qualitatively and quantitatively, comparing it to existing methods, and find that our manifold learning method outperforms state-of-the-art methods significantly. The reliability results show that more than 99% of the generated faces are acceptable as real faces.