Abstract
In this paper, we present a multi-stage regression-based approach for the 300 Videos in-the-Wild (300-VW) Challenge, which progressively initializes the shape from obvious landmarks with strong semantic meanings, e.g. eyes and mouth corners, to landmarks on face contour, eyebrows and nose bridge which have more challenging features. Compared with initialization based on mean shape and multiple random shapes, our proposed progressive initialization can very robustly handle challenging poses. It also guarantees an accurate landmark localization result and shows smooth tracking performance in real-time.