LAUSR

Recently, a local binary patterns-based initialization scheme for robust cascaded pose regression was proposed, which selects the most correlated shapes with the estimated face from training set as the initial shapes. Nevertheless, due to the massive samples in training set, the operation of selecting the most correlated shapes with the estimated shape turns out to be time-consuming. To reduce the quantity of the faces in training set for correlation analyzing in the training set, the faces should be divided into latent classes, and a face in each class could be chosen to form a smaller initial shape pool. In this paper, we view the faces and the latent classes as the latent Dirichlet allocation model. We first extract scale-invariant feature transform features from each face and employ $k$ -means algorithm on the features to obtain a fixed number of clusters. The features in each cluster are represented by its centroid, which can be used to generate the probability distribution of a face belonging to the latent classes via Gibbs sampling. The performance of the proposed scheme is evaluated on the challenging data set of Caltech Occluded Faces in the Wild. The experimental results show that the proposed scheme can significantly reduce time cost on landmark localization by 65.84% without dropping accuracy.