Design of Facial Convolutional Mesh Autoencoder
Jee-Sic Hur, Inchul Han, Hyeong-Gyun Lee, Soo Kyun Kim · 2024
In various computer vision tasks such as 3D facial reconstruction and retargeting, 3D representation poses a significant challenge. Traditional models learn linear subspaces, such as Principal Component Analysis, but they are limited in their ability to represent data beyond the input. With the advancement of deep learning, the Convolutional Mesh Autoencoder (CoMA) overcame these limitations by employing mesh simplification through Quadratic Error Metric (QEM) and utilizing the Chebyshev Network (ChebNet). However, mesh simplification based on QEM can damage feature points. In this paper, we propose Facial CoMA, which improves upon this issue by using mesh simplification through feature points and Feature Edge Quadric (FEQ).