Importance of different facial parts for face detection networks

Most state-of-the-art face detection algorithms are usually trained with full-face pictures, without any occlusions. The first novel contribution of this paper is an analysis of the accuracy of three off-the-shelf face detection algorithms (MTCNN, Retinaface, and DLIB) on occluded faces. In order to determine the importance of different facial parts, the face detection accuracy is evaluated in two settings: Firstly, we automatically modify the CFP dataset and remove different areas of each face: We overlay a grid over each face and remove one cell at a time. Similarly, we overlay a rectangle over the main landmarks of a face - eye(s), nose and mouth. Furthermore, we resemble a face mask by overlaying a rectangle starting from the bottom of the face. Secondly, we test the performance of the algorithms on people with real-world face masks. The second contribution of this paper is the discovery of a previously unknown behaviour of the widely used MTCNN face detection algorithm-if there is a face inside another face, MTCNN does not detect the larger face.