INB-Lunch-Seminar

Development of a TOF Face Detector on the ARTTS Dataset

Thomas Zielke, Jens Lippel and André Stuhlsatz

Face detection on time-of-flight image data is a challenging pattern recognition task because it involves the combination of information from two different sensor modalities. In a 2009 paper Böhme et al. show how this task can be solved by a cascade of boosted classifiers that are constructed by a modified Viola-Jones algorithm. Contrary to this approach, we formulate the detection task as a problem of dimensionality reduction. That is, an intensity image patch and its corresponding range image patch are together treated as a high dimensional feature vector that subsequently gets reduced to one dimension. Using some preprocessing operations and a conventional log-Gabor filtering on the ARTTS dataset, the size of the feature vector before dimensionality reduction amounts to 5120. The method used for dimensionality reduction has been dubbed Generalized Discriminant Analysis (GerDA). It is based on deep neural networks (DNN) that are trained in a semi-supervised fashion. Our talk covers the foundations of GerDA, the specific preprocessing steps developed for the TOF image data, the techniques to reduce the computation time of the TOF face detector, and finally the results on the ARTTS test images, including ca. 46 million non-face image patches.