Automatic 3D Facial Landmark Localization Using Models

In this research, two novel methods on automatic landmark recognition and localization on 3D face scans are developed and presented. These methods are applied to the scans after an alignment procedure that minimizes the difference between the depth values of the scan and a mean face. With the goal of producing 21 established landmarks, both Patch and Global Model methods utilize Principle Component Analysis (PCA) to model different features of the face scans. The features consisting of the models of the patches of depth maps for each landmark and the distribution of the landmarks in space are matched against the ones learned in the training phase of the Patch Model Method. In the Global Model Method, a global face model is generated by the combination of extending the model of the distribution of the landmarks in space to the whole surface of the face and modelling the residual depth values. The generated model is fitted to the scan to register the landmark locations on the scan. Evaluations of the results of the methods against the manually annotated ground truth locations of the landmarks on our database demonstrate that the methods are computationally fast and highly accurate in locating the landmarks. The methods we developed are state of the art ones with a lot of potential for further improvement