Probabilistic modelling of multiple observations in face detection Studienarbeit

The detection of human faces became a very popular field of research in computer science over the last years. Different approaches to this problem such as neuronal networks, edge-based algorithms or geometrical face models have been proposed. A very common method is to use a scanning window detector (e.g. [12], [4]). However, this typically produces a number of positive responses close by to the correct detection, which leads to the need to have a further non-maximum suppression (NMS) stage to thin out the multiple responses and to suppress spurious responses. This postprocessing is not an easy task, since it is not clear whether some detections belong to the same face or to different close (overlapping) faces and how to distinguish between true and false observations. This process is often carried out using heuristics which contain many arbitrary parameters and thresholds. In this thesis I will present a principled approach to NMS based on modelling the detections of a single object with a particular distribution (called the scale-sensitive Gaussian, SSG), and then describing the set of all detections with mixture of SSGs. I will present empirical results on the MIT+CMU faces dataset (see Appendix) showing that this approach is highly competitive with the OpenCV detector of Lienhart et al. [7].