Evaluation of Human and Machine Face Detection using a Novel Distinctive Human Appearance Dataset

Face detection is a long-standing challenge in the field of computer vision,\nwith the ultimate goal being to accurately localize human faces in an\nunconstrained environment. There are significant technical hurdles in making\nthese systems accurate due to confounding factors related to pose, image\nresolution, illumination, occlusion, and viewpoint [44]. That being said, with\nrecent developments in machine learning, face-detection systems have achieved\nextraordinary accuracy, largely built on data-driven deep-learning models [70].\nThough encouraging, a critical aspect that limits face-detection performance\nand social responsibility of deployed systems is the inherent diversity of\nhuman appearance. Every human appearance reflects something unique about a\nperson, including their heritage, identity, experiences, and visible\nmanifestations of self-expression. However, there are questions about how well\nface-detection systems perform when faced with varying face size and shape,\nskin color, body modification, and body ornamentation. Towards this goal, we\ncollected the Distinctive Human Appearance dataset, an image set that\nrepresents appearances with low frequency and that tend to be undersampled in\nface datasets. Then, we evaluated current state-of-the-art face-detection\nmodels in their ability to detect faces in these images. The evaluation results\nshow that face-detection algorithms do not generalize well to these diverse\nappearances. Evaluating and characterizing the state of current face-detection\nmodels will accelerate research and development towards creating fairer and\nmore accurate face-detection systems.\n