An Asian Face Dataset and How Race Influences Face Recognition

The face recognition scheme based on deep learning can give the best face recognition performance at present, but this scheme requires a large amount of labeled face data. The currently available large-scale face datasets are mainly Westerners, only containing few Asians. In practice, we have found that models trained using these data sets are lower in accuracy in identifying Asians than Westerners. Therefore, the establishment of a large-scale Asian face dataset is of great value for the development and deployment of face related applications for Asians. In this paper, we propose a simple semi-automatic approach to collect face images from Internet and build a large-scale Asian face dataset (AFD) containing 2019 subjects and 360,000 images. To the best of our knowledge, this is the largest Asian face image dataset proposed so far. To illustrate the quality of AFD, we train 3 different models with the same CNN structure yet by different training datasets (AFD, WebFace, mixed WebFace&AFD) and verify them on one Western and two Asian face testing datasets. Extensive experimental results show that the model by our AFD outperforms counterparts by a large margin for Asian face recognition. We have made the AFD dataset public to facilitate face recognition development for Asians.