Face Alignment Assisted by Head Pose Estimation

摘要: In this paper we propose a supervised initialization scheme for cascaded face alignment based on explicit head pose estimation. We first investigate the failure cases of most state art approaches and observe that these failures often share one common global property, i.e. variation is usually large. Inspired by this, deep convolutional network model reliable accurate Instead using mean shape, or randomly selected shapes initialisation, two schemes generating initialisation: relies projecting 3D shape (represented facial landmarks) onto 2D image under estimated pose; second searches nearest neighbour from training set according to distance. By doing so, initialisation gets closer actual which enhances possibility convergence in turn improves performance. demonstrate proposed method benchmark 300W dataset show very competitive performance both estimation alignment.