Detecting Faces With Face Masks

This paper deals with the evaluation of several methods for face detection when the face is covered by a mask. The methods evaluated are Haar cascade and Histogram of Oriented Gradients as feature-based approaches, Multitask Cascade Convolutional Neural Network, Max Margin Object Detection and TinyFace as convolutional neural network based approaches. Various types of face masks are considered: disposal face mask, burka, balaclava, ski helmet with ski goggles, hockey helmet with protective grill, costumes, and others. The TinyFace method achieves the best accuracy result, but also requires much more computational power than other approaches. Therefore, this paper describes an experiment to see if the accuracy of some of the remaining methods can be improved by retraining their models with new image data containing faces with various face masks.