Haar Cascade Face Detector Quality Dependence on Training Dataset Variablity

Abstract

Background. When training generalized face detectors based on Haar cascades, there is a problem of long learning time of the resulting cascades and their poor quality. Therefore, in practice, frontal and profile face detectors are trained separately. Such approach makes recognition systems more complex.

Objective. The aim of the paper to compare the impact of the training set composition with faces at different inclination angles on the quality of the trained detectors.

Methods. It is proposed to train a series of face detectors on sub-sets that cover different ranges of face angles. All other parameters of training are fixed. As the result, the learning time and the quality of the obtained cascades will be compared.

Results. The quality and the training time of face classifiers are evaluated depending on the composition of the training subsets. Also the quality of the frontal and side face classifiers is compared having the same sizes of training sets. It is shown that the AUC metric has a difference of 0.003 between the frontal and profile face detectors.

Conclusions. It has been shown experimentally that the more variations present in the object’s dataset (the side-view of faces compared to the frontal positions), the longer and harder the Haar cascade learns, given the same amounts of the training samples. Using the proposed approach, the quality of the final classifier can be controlled by selecting the appropriate composition of the training samples.