Iterative Face Detection from the Global to Local

To balance between accuracy and speed is a dilemma in face detection, and scale variance problem is one of the main causes. In this paper we propose fast and accurate iterative face detection method processing from the global to local. We define a class of object called “probable regions” which contain small faces. “Probable regions” are detected iteratively in order to enlarge the small face parts. Thus small faces turn into large faces after several iterations. We design a strategy of training samples augmentation to meet the requirement for two object classes, so that extra annotation is unneeded. Our method is simple and clear to deploy. Experiments show that our method achieves competitive accuracy with real time speed. Detection time consumption will not explicitly grow when resolution of sample image increases. The speed is merely related to actual amount of faces, which adapts to real world applications.