Spatial attention network for head detection

Abstract

Human head detection is widely used in computer vision. However, in practical applications, human head detection is likely to cause false alarms because of the angle, light condition, and cameras. This paper proposes a novel spatial attention network (SAN) which adopts the saliency module to exploit the environmental information beyond the proposal which is ignored in the Faster-RCNN. At the meantime, the class score and saliency score are fused together through a suitable strategy to effectively suppress false positive samples. In order to train and test our model, this paper has established a dataset including 55,802 images. We have evaluated our method and the final experimental results show that our model is significantly superior to the Faster-RCNN model.