The geometry of a face can tell many things about someone. It harbours information on gender, facial expression, identity and even birth defects. Many contemporary facial expression recognition systems employ state of the art trackers, that track the positions of facial points throughout a video. However, how to initialise these trackers accurately and robustly is still ongoing research.
One of our aims is to perform a fast and reliable location of these facial points in order to perform further face analysis, using both visible and thermal imagery. To this end, we apply Machine Learning techniques (SVRs and boosting) in the Computer Vision field. Our initial efforts to this end resulted in a reliable facial point detector based on the responses of Gabor filters and ensemble learning (Gntleboost). This work is explained in . The compailed version of the detector can be found in the 'Resources'. Our current efforts aim at obtaining a reliable real-time detection, so that the software can be used in real-life applications. Specifically, we use regression techniques to obtain fast estimates of the facial feature location, and combine the predictions of multiple points into probabilistic models that encode the geometric configuration of the face to constrain the solutions and boost the robustness of the method. More information about this research can be found in .
A complied version of this work will be available from this website soon for Mac and Windows.
M. Pantic, M. Tomc, L. Rothkrantz. Proceedings of IEEE Int'l Conf. Systems, Man and Cybernetics (SMC'01). Tucson, USA, pp. 1188 - 1193, October 2001.