Sampling strategies for performance improvement in cascaded face regression

Automatic face landmarking has received a lot of attention in the past decades. It is now mature enough to be implemented in fully autonomous video systems. As cascade-of-regression based algorithms have become state of the art in such systems, two major (and still relevant) sources of interest have slowly faded away: the need for semantic-driven learning beyond ground truth annotation, and full video chain performance i.e. tracking efficiency, which in the case of said methods strongly relates to their robustness towards shape initialization before fitting. In this paper, we investigate how data sampling using face priors can affect their performance in terms of convergence and robustness. We propose new strategies based on said priors to overcome inconsistencies observed during cascade-of-regression learning on purely random sampling-based stages. We will show that simple choices can be easily integrated within regression-based face tracking systems to increase accuracy and robustness.