Estimation of Unmasked Face Images Based on Voice and 3DMM

Facemasks have become common due to the COVID-19 pandemic. They have begun to affect security and identification systems because they cover almost half of the face. Current state-of-the-art methods have been applied to estimate unmasked faces from masked face images. They are successful in improving the quality of the face texture by 3D Morphable Model (3DMM) as intermediate representations. However, their performance in restoring the face shapes is insufficient, and some of generated faces lack identities. In this study, we focus on voice, which has a particularly high correlation with the shape of the mouth and nose, which are obscured by masks. We propose a multimodal method using 3DMM and voice for face shape estimation under masks. Experimental results show that the proposed method qualitatively and quantitatively improves the quality of shape restoration of a face compared to the baseline method without considering voice.