Fine-grained evaluation on face detection in the wild

Current evaluation datasets for face detection, which is of great value in real-world applications, are still somewhat out-of-date. We propose a new face detection dataset MALF (short for Multi-Attribute Labelled Faces), which contains 5,250 images collected from the Internet and ~12,000 labelled faces. The MALF dataset highlights in two main features: 1) It is the largest dataset for evaluation of face detection in the wild, and the annotation of multiple facial attributes makes it possible for fine-grained performance analysis. 2) To reveal the `true´ performances of algorithms in practice, MALF adopts an evaluation metric that puts stress on the recall rate at a relatively low false alarm rate. Besides providing a large dataset for face detection evaluation, this paper also collects more than 20 state-of-the-art algorithms, both from academia and industry, and conducts a fine-grained comparative evaluation of these algorithms, which can be considered as a summary of past advances made in face detection. The dataset and up-to-date results of the evaluation can be found at http: //www.cbsr.ia.ac.cn/faceevaluation/.