Attention-guided Face Alignment Based on Inverted Residuals
Xinjie Tong, Teng Wang · 2019
In this paper, we build an attention-guided network based on inverted residuals for face alignment. This network consists of a multi-resolution branches structure module and an attention-guided feature fusion module. These two modules employs inverted residuals to construct a light weight network. The multi-resolution branches structure consists of two branches, high-resolution branch and low-resolution branch. The high-resolution branch extracts low-level features from the original faces; The low-resolution branch generates the attention maps by using low-level features from the trunk. The attention-guided feature fusion module is designed to merge features from two branches based on attention mechanism. Extensive experiments are conducted on two public benchmark datasets, 300-W and AFLW. We compare our method with the state-of-the-art methods to indicate the superiority of our method. We achieve 3.49%mean error on 300-W full set, and 1.47%mean error on AFLW-Full datasets. Meanwhile, our method achieves the real-time detection (10ms) on 300-W with 68 landmarks.