Towards robust cascaded regression for face alignment in the wild

Most state-of-the-art solutions for localizing facial feature landmarks build on the recent success of the cascaded regression framework [7, 15, 34], which progressively predicts the shape update based on the previous shape estimate and its feature calculation. We revisit several core aspects of this framework and show that proper selection of regression method, local image feature and fine-tuning of further fitting strategies can achieve top performance for face alignment using the generic cascaded regression algorithm. In particular, our strongest model features Iteratively Reweighted Least Squares (IRLS) [18] for training robust regressors in the presence of outliers in the training data, RootSIFT [2] as the image patch descriptor that replaces the original Euclidean distance in SIFT [24] with the Hellinger distance, as well as coarse-to-fine fitting and in-plane pose normalization during shape update. We show the benefit of each proposed improvement by extensive individual experiments compared to the baseline approach [34] on the LFPW dataset [4]. On the currently most challenging 300-W dataset [28] and COFW dataset [4], we report state-of-the-art results that are superior to or on par with recently published algorithms.