Abstract
This thesis studies robust facial landmark detection (FLD) algorithms for faces in the wild. In general, a facial landmark detector is applied to a face bounding box, generated by a face detector, to obtain accurate geometric positions of a set of predefined landmarks for a face image. The landmarks are usually needed to extract accurate facial texture features in subsequent stages in an automatic face analysis system. Unfortunately, in uncontrolled scenarios, the variations in appearance caused by pose, expression, illumination and occlusion pose obstacles and difficulties in FLD. Classical algorithms often fail for faces in the wild in the presence of these variations. To meet the requirements for robust FLD in the wild, the thesis presents three main contributions to the field: Firstly, to achieve variation-invariant FLD, we study the tensor-based active appearance model. One of the difficulties of using a tensor model is that it requires a complete training dataset that includes training samples of all modes of variation for each subject, but in practice we often encounter the problem of missing training samples. To deal with this issue, we propose the use of tensor completion methods to reconstruct missing shape and global texture in tensor-based active appearance model. Secondly, in recent years, discriminative cascaded regression has received extensive attention in FLD. We consider the problem of scale variation in shape update and local feature extraction when using a regression-based model, and develop an adaptive scheme for scale-invariant FLD. In addition, a new random cascaded regression copse structure has been designed to improve the generalization capability of the cascaded regression method. Lastly, since cascaded regression is supervised, a large amount of training data is crucial. However, the task of providing training samples is often time-consuming, involving a considerable amount of tedious manual work. As an alternative, we propose the use of a 3D morphable face model to generate synthesised faces for regression-based detector training. To adapt the model trained on the synthetic data to real face images, we propose a cascaded collaborative regression algorithm. The training is based on a mix of synthetic and real image data with the mixing controlled by a dynamic mixture weighting schedule.