L 2, 1-based regression and prediction accumulation across views for robust facial landmark detection

We propose a new methodology for facial landmark detection. Similar to other state-of-the-art methods, we rely on the use of cascaded regression to perform inference, and we use a feature representation that results from concatenating 66 HOG descriptors, one per landmark. However, we propose a novel regression method that substitutes the commonly used Least Squares regressor. This new method makes use of the L2,1 norm, and it is designed to increase the robust- ness of the regressor to poor initialisations (e.g., due to large out of plane head poses) or partial occlusions. Furthermore, we propose to use multiple initialisations, consisting of both spatial translation and 4 head poses corresponding to different pan rotations. These estimates are aggregated into a single prediction in a robust manner. Both strategies are designed to improve the convergence behaviour of the algorithm, so that it can cope with the challenges of in-the- wild data. We further detail some important experimental details, and show extensive performance comparisons highlighting the performance improvement attained by the method proposed here