Generic face alignment using an improved Active Shape Model

Although conventional Active Shape Model (ASM) and Active Appearance Model (AAM) based approaches have achieved some success, however, evidence suggests that the performance of a person-specific face alignment which aligns the variation in appearance of a single person across pose, illumination, and expression is substantially better than the performance of generic face alignment which aligns the variation in appearance of many faces, including unseen faces not in the training set. This paper proposes a discriminative framework for generic face alignment. This technique is presented under the framework of conventional Active Shape Model (ASM) but has three improvements. First, random forest classifiers are trained to recognize local appearance around each landmark. This discriminative learning provides more robustness weight for the optimization fitting procedure. Second, to impose constrains, shape vectors are restricted to the vector space spanned by the training database. Third, data augment scheme is used for the benefit of a large training set. Experimental results show that this approach can achieve good performance on generic face alignment.