An end-to-end generative adversarial network for crowd counting under complicated scenes

Crowd counting and analyzing its distribution is a challenging video surveillance application. In this paper, a totally novel, end-to-end way to estimate the crowd number under complicated scenes is proposed. For the purpose, we apply the conditional adversarial networks to translate the input image to its density map. The conditional generative adversarial model is trained with input image and its corresponding density image. The proposed method avoid the design of complex CNN architecture to extract specific property features. Besides, no more data augmentation is needed in our method. Evaluated on the dataset of Shanghaitech which consists of two challenge parts, our methods shows convincing counting results with high quality estimated density images. Moreover, our experiments can been done in an efficient and labor saving way.