Generating Facial Ground Truth with Synthetic Faces

This work describes a methodology for generation of facial ground truth with synthetic faces. Our focus is to provide a way to generate accurate data for the evaluation of Computer Vision algorithms, in terms of facial detection and its components. Such algorithms play a key role in face detection. We present a prototype in which we can generate facial animation videos using a 3D face models database, controlling face actions, illumination conditions and camera position. The facial animation platform allows us to generate animations with speech, facial expressions and eye motion, in order to approach realistic human face behavior. In addition, our model provides the ground truth of a set of facial feature points at each frame. As result, we are able to build a video database of synthetic human faces with ground truth, which can be used for training/evaluation of several algorithms for tracking and/or detection. We also present experiments using our generated videos to evaluate face, eye and mouth detection algorithms, comparing their performance with real video sequences.