Multi-angle Face Detection with Step-by-Step Adjustment Networks

Multi-angle face detection, has a wide range of application needs in our lives. However, due to the large difference in facial features at different angles, it brings certain challenges for detection. In order to solve this problem more efficiently, we propose Step-by-Step Adjustment Networks(SAN) to perform multi-angle face detection in a coarse-to-fine manner. The method contains three stages, the first stage is that distinguish the target is human or cat's face or non-face. Then each candidate target's face orientation will be adjust to upright gradually. Initially, the position of bounding box need to be divided by several steps, and only predicting coarse orientations. Our SAN can achieve gradually decreasing the rotate ranges, and obtain accurately detect consequence with full 360? rotate angles. The experiments on Multi-Oriented FDDB and a challenging subset of WIDER FACE containing rotated facial include human and cat's, the result show that our SAN can achieve good performance.