Face Detection for Crowd Analysis Using Deep Convolutional Neural Networks

Crowd analysis is a challenging topic within computer vision, current state of the art methods for face detection in crowds suffer from poor results due to visual occlusions, scene semantics and overlapping subjects. In this work, we propose a novel approach of utilizing existing semantic segmentation methods to detect and segment faces in obscured images. We use an implementation of Mask RCNN trained on the popular Labelled Faces in the Wild (LFW) database to compare performance with Viola Jones, histogram of orientated gradients and max-margin object detection using a synthetically generated occluded subset of LFW. Results show that when images contain fair sized occlusions, Mask RCNN outperforms the current state of the art method. State of the art performance was achieved on this dataset and context specific improvements are suggested for further work. The contribution of this paper is not to regurgitate the finding from the original paper on Mask RCNN but provide results on the efficiency of using the method in the context of face detection for crowd analysis. Additionally, exploration of suitable hyper parameters for this context has been performed and described. Code has been made publicly available.