3D Face Reconstruction Based on Improved CANDIDE-3 Model

Abstract

In this paper, an individual 3D face reconstruction method is proposed. A 3D facial model is constructed from a static or captured image from video camera, by procedures like face region detection, feature points detection, blink detection, feature points matching and 3D facial texture mapping. In order to improve performance of reconstruction, a pattern-based blink detection method is proposed in this paper. During the reconstruction, once blink is detected, further computing will be skipped. CANDIDE-3 model, which is just a parameterization grid, lacks ability to determine blink, blink detection is significant. Compared with blink detection base on Otsu, the blink detection method in this paper has higher performance and stronger flexibility toward various illumination conditions. Experimental results show that there are good results of 3D face reconstruction.