Automatic landmarking for non-cooperative 3D face recognition

This thesis describes a new framework for 3D surface landmarking and evaluates its performance for feature localisation on human faces. This framework has two main parts that can be designed and optimised independently. The first one is a keypoint detection system that returns positions of interest for a given mesh surface by using a learnt dictionary of local shapes. The second one is a labelling system, using model fitting approaches that establish a one-to-one correspondence between the set of unlabelled input points and a learnt representation of the class of object to detect. Our keypoint detection system returns local maxima over score maps that are generated from an arbitrarily large set of local shape descriptors. The distributions of these descriptors (scalars or histograms) are learnt for known landmark positions on a training dataset in order to generate a model. The similarity between the input descriptor value for a given vertex and a model shape is used as a descriptor-related score. Our labelling system can make use of both hypergraph matching techniques and rigid registration techniques to reduce the ambiguity attached to unlabelled input keypoints for which a list of model landmark candidates have been seeded. The soft matching techniques use multi-attributed hyperedges to reduce ambiguity, while the registration techniques use scale-adapted rigid transformation computed from 3 or more points in order to obtain oneto-one correspondences. Our final system achieves better or comparable (depending on the metric) results than the state-of-the-art while being more generic. It does not require pre-processing such as cropping, spike removal and hole filling and is more robust to occlusion of salient local regions, such as those near the nose tip and inner eye corners. It is also fully pose invariant and can be used with kinds of objects other than faces, provided that labelled training data is available.