Salient-points-guided face alignment

Abstract

Regression-based face alignment approach is fast and accurate but is always limited by the initial face. Aim at limitation of initialization in regression methods, this paper presents a novel two-stage framework named salient-points-guided face alignment. In first stage, we use cascade regression framework to train a salient points (eye centers, nose, mouth corners) localization model. Then the salient points information is used as a guidance for searching the similar faces from training set. In second stage, leveraging the similar faces to generate the initial face for all points regression. In order to give more comprehensive comparison, a new evaluation metric is proposed. Considering the global distance between estimated face and ground-truth, the new evaluation metric is defined as sum of the global distance and the widely used average point-to-point distance. The results show that our approach can achieve state-of-the-art performance (12% higher than the human performance on COFW) and the new evaluation metric is more reasonable.