Adaptive Thresholds for Robust Face Detection with a Short Cascade of Classifiers

In this paper we present the preliminary results of our research on face detection which show that it is possible to significantly improve the performance of a single strong classifier (in terms of high detection rate and reduced number of false positives) by adapting the detection threshold to each input image instead of using the fixed thresholds learned in the training step. Moreover, if this adaptive thresholds are used at the last stage of a short cascade of classifiers (less than 5 stages in all), we show that the performance of the cascade is close to that of a longer ´classical´ cascade, while its computational cost is much lower.