Analysis of Self-Occlusion in Single View-based 3D Face Reconstruction
Youn Joo Lee, Sung Joo Lee, Jaeik Jo, Kang Ryoung Park, Jaihie Kim · ICEIC : International Conference on Electronics, Informations and Communications · 2010
3D face reconstruction based on a single 2D image has been actively researched in computer vision because it has many useful applications such as 3D faces recognition, 3D face generation of game character, and so on. Most of previous methods require an assumption of a frontal pose, but this assumption cannot be always satisfied in real world application. To overcome this problem, renowned 3D morphable model (3DMM)-based methods reconstruct a 3D facial shape without the constraint of a frontal pose. However, these methods could be sensitive to the self-occlusion errors caused by pose variations. In this paper, we analyze the self-occlusion problem in a simplified 3DMM-based 3D face reconstruction and proved that the problem leads to performance degradation of 3D face reconstruction in the experimental results.