Learning good features for Active Shape Models

Active Shape Models (ASMs) are commonly used to model the appearance and shape variation of objects in images. This paper proposes two strategies to improve speed and accuracy in ASMs fitting. First, we define a new criterion to select landmarks that have good generalization properties. Second, for each landmark we learn a subspace with improved facial feature response effectively avoiding local minima in the ASM fitting. Experimental results show the effectiveness and robustness of the approach.