Learning deep compact channel features for object detection in traffic scenes.