Face detection is challenging as faces in images could be present atarbitrary locations and in different scales. We propose a three-stage cascadestructure based on fully convolutional neural networks (FCNs). It firstproposes the approximate locations where the faces may be, then aims to findthe accurate location by zooming on to the faces. Each level of the FCN cascadeis a multi-scale fully-convolutional network, which generates scores atdifferent locations and in different scales. A score map is generated aftereach FCN stage. Probable regions of face are selected and fed to the nextstage. The number of proposals is decreased after each level, and the areas ofregions are decreased to more precisely fit the face. Compared to passingproposals directly between stages, passing probable regions can decrease thenumber of proposals and reduce the cases where first stage doesn't propose goodbounding boxes. We show that by using FCN and score map, the FCN cascade facedetector can achieve strong performance on public datasets.
展开▼
