In this context, we propose a fully automatic multilinear algorithm that can estab-lish dense correspondence (≥ 13, 500 points) over a large number of 3D human faceswith varying identities and expressions. We first train a deep Convolutional NeuralNetwork (CNN) for landmark identification. CNNs have been extensively used for 2Dtexture images where the training data with ground truth labels is abundantly available.However, in case of 3D faces there is a dearth of training data that contains significantvariation in facial shape, ethnicity and expressions. Our Deep Landmark IdentificationNetwork (DLIN) is trained on synthetic 3D images generated from a commercial soft-ware (FaceGenTM) and is able to detect 11 biologically significant [28] facial landmarkswith high accuracy and efficiency. Next, we divide the 3D face into five Voronoi regionsaround a subset of these landmarks using geodesically evolved level set curves [29]. Asparse set of discriminative keypoints are detected within each region and used to elas-tically align the corresponding region shapes of two given faces. Dense correspondenceis then achieved through nearest neighbour matches between the region vertices of alltraining faces. Finally, a 3D deformable model is constructed from the densely corre-sponding faces and the correspondence information is transferred to unseen 3D facesby fitting the deformable model in an iterative optimization.

Face alignment, which fits a face model to an imageand extracts the semantic meanings of facial pixels, hasbeen an important topic in CV community. However , mostalgorithms are designed for faces in small to medium pos-
es (below 45◦), lacking the ability to align faces in largeposes up to 90◦. The challenges are three-fold: Firstly, thecommonly used landmark-based face model assumes thatall the landmarks are visible and is therefore not suitablefor profile views. Secondly, the face appearance variesmore dramatically across large poses, ranging from frontalview to profile view. Thirdly, labelling landmarks in largeposes is extremely challenging since the invisible landmarkshave to be guessed. In this paper , we propose a solution tothe three problems in an new alignment framework, called3D Dense Face Alignment (3DDF A), in which a dense 3Dface model is fitted to the image via convolutional neutralnetwork (CNN). We also propose a method to synthesizelarge-scale training samples in profile views to solve thethird problem of data labelling. Experiments on the chal-lenging AFLW database show that our approach achievessignificant improvements over state-of-the-art methods.