An Efficient Training Strategy for Face Detector in Specific Scenes

摘要

Face detection has been well studied and widely applied in a variety of fields including online education, computer-aided medicine, and video surveillance, etc. Unfortunately, directly applying the algorithm trained on public wild face benchmarks to unconstrained scenes fails to obtain satisfactory performance. To solve this problem, we first propose an automatic data annotation method, and then propose a simple but efficient self-adapted training strategy for the face detector based on aggregate channel features and the boosting classifier. Experiments show that the self-adapted detector outperforms several other state-of-the-art approaches on our challenging test set.