Our approach benefits from the recent advances in cascade face alignment [19,
4, 28, 29, 21]. In such works, the face shape is progressively updated via boosted
regression. The regression learning in each stage not only depends on the image,
but also depends on the estimated shape from the previous stage. Features learnt
this way are called shape indexed features. Such features present more invariance
to the geometric variations in the face shapes and they are crucial for high alignment
accuracy and speed. As the cascade structure has been proven effective in
detection and alignment, we propose to combine the two to benefit each other. In
Section 3, we present a general cascade framework that unifies the two tasks. The
detection learning is made more effective with embedded alignment information,
and alignment is achieved simultaneously without performing it separately. In
Section 4, we extend the recent state-of-the-art alignment method [21] under the new framework