Analyzing the Effect of Eye Center Localization on Accurate Landmark Localization in a Facial Image

摘要

Localization of facial landmarks on a human face is an important step for many face-related computer vision applications. The most of the earlier techniques (AAMs, CLMs) has achieved good performance in landmark localization but they always limited by the initialization of landmarks. In this paper, the initialization problem is solved by taking the eye center as references to the mean face shape. In the proposed method, the eye centers are estimated using multi-scale iris shape feature first and then the constrained local model is applied for landmark localization where initialization is done using mean face shape taking eye centers as references. The performance of eye center estimation and landmark localization method are evaluated on AR and Multi-PIE databases. For eye center estimation three normalized eye localization error is considered whereas for landmark localization RMSE and detection rate are considered. For landmark localization, a total of 130 and 68 landmarks are considered for AR and Multi-PIE database respectively. The experimental results suggest that the proposed method has achieved improved performance as compared to some of the other methods.