Design of a Deep Face Detector by Mask R-CNN

Abstract

In this work an existing object detector, Mask RCNN, is trained for face detection and performance results are reported by using the learned model. Differing from the existing work, it is aimed to train the deep detector with a small number of training examples and also to perform instance segmentation along with an object bounding box detection. Training set includes 2695 face examples collected from PASCAL-VOC database. Performance has been reported on 159,000 test faces of WIDER FACE benchmarking database. Numerical results demonstrate that the trained Mask R-CNN provides higher detection rates with respect to the baseline detector [1], particularly 6%, 12%, and 3% higher face detection accuracy for the small, medium and large scale faces, respectively. It is also reported that our performance outperforms Viola & Jones face detector. We released the face segmentation ground-truth data that was used to train Mask R-CNN and training-test routines developed in TensorFlow platform to public usage at our GitHub repository.