A Rotation-Invariance Face Detector Based on RetinaNet

Abstract The use of deep convolutional neural networks has greatly improved the performance general face detection. For detecting rotated faces, mainstream approach is to multi-stage detectors gradually adjust a vertical orientation for detection, which increases complexity training as multiple are involved. In this study, we propose new method rotation-invariant abandons previously used cascaded architecture with stages and instead uses single-stage detector achieve end-to-end detection classification, box regression, facial landmark regression. Extensive experiments on FDDB in orientations have shown effectiveness our method. results demonstrate that achieves good accuracy even exceeds other front-facing dataset.