Face Detection in Challenging Scenes with a Customized Backbone

Existing works in deep network-based face detection have mostly used pre-trained backbone networks like ResNet-50/101/152 and VGG16/19. These networks are trained on large image collections and can produce rich image representations. Only a few works have attempted the development of customized backbones for the face detection task. The first contribution of this work proposes a novel backbone network with 7.5M parameters and 43 GFLOPs. The overall face detector is motivated by the RetinaFace architecture. The second contribution proposes a detector head considering the imbalance in face and non-face. The proposed face detector is benchmarked on the WIDER FACE and FDDB dataset and is observed to provide competitive performance against state-of-art approaches.