Structure-aware Heatmap and Boundary Map Regression Based Robust Face Alignment

Abstract

Large head pose variations and severe occlusion are challenging problems for face alignment. In this paper, we propose a Structure-aware Heatmap and Boundary map Regression Network (SHBRN), consisting of a rough estimation network and a refinement network, to accounting for the structural geometry of faces via the boundary map. Specifically, in the rough estimation network, a structure-aware module is designed to capture low-level features rich in structure information, and both heatmaps and boundary maps are predicted by the hourglass network. In this way, the network can not only estimate the initial location of keypoints, but also implicitly take the geometric structure into consideration. In the refinement network, the boundary maps and heatmaps are fused with the features extracted in the rough stage via attention mechanism. As a result, the network can combine the global information with local appearance for obtaining complete face representations, and also optimize the spatial relationship of different keypoints. Our proposed network is superior to the existing methods on 300W, COFW, and AFLW datasets, especially for those challenging situations, which proves the effectiveness and robustness of our model.

Keywords

• distance learning
• image analysis
• neural network
• pattern analysis
• supervised learning