Floating-point precision and deformation awareness for scalable and robust 3D face alignment

Abstract

This paper improves the accuracy of heatmap-based 3D face alignment neural networks. Many current approaches in face alignment are limited by two major problems, quantization and the lack of regularization of heatmaps. The first limitation is caused by the non-differentiable ar?max function, which extracts landmark coordinates from heatmaps as integer indices. Heatmaps are generated at low-resolution to reduce the memory and computational costs, which results in heatmaps far lower than the input image's resolution. We propose a heatmap generator network producing floatingpoint precision heatmaps that are scalable to higher-resolutions. To resolve the second limitation, we propose a novel deformation constraint on heatmaps. The constraint is based on graph-Laplacian and enables a heatmap generator to regularize overall shape of the output face landmarks using the global face structure. By eliminating quantization and including regularization, our method can vastly improve landmark localization accuracy, and achieves the state-ofthe- art performance without adding complex network structures.