A two-stage estimation method for depth estimation of facial landmarks
Xun Gong, Zehua Fu, Xinxin Li, Lin Feng · 2015
To address the problem of 3D face modeling based on a set of landmarks on images, the traditional feature-based morphable model, using face class-specific information, makes direct use of these 2D points to infer a dense 3D face surface. However, the unknown depth of landmarks degrades accuracy considerably. A promising solution is to predict the depth of landmarks at first. Bases on this idea, a two-stage estimation method is proposed to compute the depth value of landmarks from two images. And then, the estimated 3D landmarks are applied to a deformation algorithm to make a precise 3D dense facial shape. Test results on synthesized images with known ground-truth show that the proposed two-stage estimation method can obtain landmarks’ depth both effectively and efficiently, and further that the reconstructed accuracy is greatly enhanced with the estimated 3D landmarks. Reconstruction results of real-world photos are rather realistic.