Stacked Multi-Target Network for Robust Facial Landmark Localisation.

Yun Yang, Bing Yu, Xiaodong Li, Bailan Feng

Abstract: We thoroughly analyse regression-based face alignment methods and introduce a novel stacked multi-target network for robust facial landmark localisation. The primary heatmap regression-based network concentrates on locating the coarse position of pre-defined landmarks while the secondary coordinate regression-based network is responsible for modelling fine sub-pixel features. Specifically, we elaborate the differences among widely-used Cross Entropy related loss functions and propose a new Bilateral Inhibition Cross Entropy loss function, which enlarges the margin between elements in the output heatmaps. Besides, in order to deal with the discrepancy between optimization and evaluation, we propose to dynamically adjust the radius of kernel function during the training process. We demonstrate that training with decreasing radius in temporal order performs much better than assigning it spatially, i.e. decreasing radius along the stages of stacked hourglass networks. Finally, we innovatively limit the output of the secondary coordinate regression network to a reasonable range by importing the hinge loss to refine the coarse coordinate locations for sub-pixel accuracy. Extensive experiments on public datasets such as 300-W, COFW, and AFLW demonstrate that our proposed method performs superiorly to the state-of-the-art approaches.