Multi-Scale Fully Convolutional Network for Fast Face Detection
Yancheng Bai, Wenjing Ma, Yucheng Li, Liangliang Cao, Wen Guo and Luwei Yang
Image pyramid is a common strategy in detecting objects with different scales in an image. The computation of features at every scale of a finely-sampled image pyramid is the computational bottleneck of many modern face detectors. To deal with this problem, we propose a multi-scale fully convolutional network framework for face detection. In our detector, face models at different scales are trained end-to-end and they share the same convolutional feature maps. During testing, only images at octave-spaced scale intervals need to be processed by our detector. And faces of different scales between two consecutive octaves can be detected by multi-scale models in our system. This makes our detector very efficient and can run about 100 FPS on a GPU for VGA images. Meanwhile, our detector shows superior performance over most of state-of-the-art ones on three challenging benchmarks, including FDDB, AFW, and PASCAL faces.