Metric Learning Based False Positives Filtering for Face Detection

Nanhai Zhang, Jiajie Han, Jiani Hu, Weihong Deng

Abstract: Face detection in the wild is a challenging task within the field of computer vision. Many face detectors fail to distinguish face images and non-face images because intra-class variations surpass inter-class variations. To overcome it, we propose a metric learning based false positives filtering for face detection. With 8 average faces as standard face, we apply metric learning to seek a linear transformation to reduce the distance between face images and standard faces while enlarge the distance between non-face images and standard faces. To solve our defining objective function for metric learning, we adopt a batch-stochastic gradient descent scheme, with which we can get stable solution fast. The results on FDDB and our self-collected dataset show a good performance of our method for improving Viola-Jones face detectors.