Exemplar-based Cascaded Stacked Auto-Encoder Networks for robust face alignment

In this paper, we present a novel Exemplar-based Cascaded Stacked Auto-Encoder Network (ECSAN) for facial landmarks detection. The proposed framework consists of a Global Exemplar Constraint Stacked Auto-Encoder Network (GECSAN) and a set of Local Information Preserve Stacked Auto-Encoder Networks (LIPSANs). In our work, GECSAN utilizes successive stacked auto-encoder network and some well-designed exemplars to obtain an initial shape estimation from a holistic facial image. Then LIPSANs are presented which take the local features extracted around current landmarks as input and generate a facial landmark refinement. Different from existing deep models, a prior exemplar-based shape is utilized to handle the partial occlusion in the image so that our model can achieve robustness against local occlusions. Experimental results on several datasets demonstrate that our model acquires better performance over the state-of-the-art methods with respect to occlusion handling and attain higher alignment accuracy.