ABSTRACT
Facial key-point localization refers to detecting salient facial landmarks on the human face. Current state-of-the-art approaches for facial key-point localization are based on cascaded regression. This chapter presents the related work in deep cascaded regression. To tackle the aforementioned challenges in deep cascaded regression models, the authors propose a globally optimized dual-pathway (GoDP) architecture where all inferences are conducted on 2D score maps to facilitate gradient back-propagation. Because there are very few landmark locations activated on the 2D score maps, a distance-aware softmax function (DSL) that reduces the false alarms in the 2D score maps is proposed. To solve the spatial-semantic uncertainty problem of deep architecture, a dual-pathway model where shallow and deep layers of the network are jointly forced to maximize the possibility of highly specific candidate regions is proposed. As a result, facial key-points localization model achieved state-of-the-art performance on multiple challenging databases.
