Depth-wise and point-wise convolutional layers for the backbone.

Performance on WIDER Face: AP_easy=0.856, AP_medium=0.842, AP_hard=0.727