Occluded and Tiny Face Detection with Deep and Shallow Features Fusion and Compensation

Face detection has been widely developed in the past decades. However, detecting occluded and tiny faces still remains great challenges. Previous methods solve these problems to a certain extent, but they still have two main shortcomings. Firstly, the feature fusion has positive significance for multi-scale detection, but since the features in deep and shallow layers are essentially different, blindly feature fusion from different layers may lead to the interferences. Secondly, the interrelation between the background and the faces is easy to be neglected, so that some features belonging to faces may be misclassified to background. In order to solve the first problem, we present two special feature pyramid networks to adaptively integrate the deep and shallow features, respectively, named SFPN and DFPN. SFPN focuses on the fusion of the shallow features while DFPN concentrates on the deep features fusion. As for the second problem, we come up with a module named background-assisted compensation module (BACM), which can enhance the interrelation between the background and the faces and compensate for the face features that are classified into the background mistakenly. Our detector has achieved superior performance compared with other corresponding methods on wider face dataset.