Localizing Facial Keypoints with Global Descriptor Search, Neighbour Alignment and Locally Linear Models

We present our technique for facial key point localization in the wild submitted to the 300-W challenge. Our approach begins with a nearest neighbour search using global descriptors. We then employ an alignment of local neighbours and dynamically fit a locally linear model to the global key point configurations of the returned neighbours. Neighbours are also used to define restricted areas of the input image in which we apply local discriminative classifiers. We then employ an energy function based minimization approach to combine local classifier predictions with the dynamically estimated joint key point configuration model. % Our method is able place 68 key points on in the wild facial imagery with an average localization error of less than 10% of the inter-ocular distance for almost 50% of the challenge test examples. Our model therein increased the yield of low error images over the baseline AAM result provided by the challenge organizers by a factor of 2.2 for the 68 key point challenge. Our method improves the 51 key point baseline result by a factor of 1.95, yielding key points for more than 50% of the test examples with error of less than 10% of inter-ocular distance.