Crowd Counting via Conditional Generative Adversarial Networks

Most of current crowd counting algorithms use Euclidean loss to narrow the gap between density map and ground-truth, which leads to the low quality of density maps. In order to alleviate the above problems, we propose a crowd counting method based on conditional generative adversarial framework, which utilizes the game between generator and discriminator to achieve high quality conversion of crowd images to density maps. Specifically, CSRNet is designed as a generator, which uses the method of dilated convolution to extract the detailed information of images under the condition of adapting the scale variation. PatchGAN is designed as a discriminator to simulate high-frequency structures to further improve the quality of density maps. Benefiting from the joint optimization of adversarial loss and L2 loss, our framework can not only accurately capture the low-frequency informations, but also better model the high-frequency informations. We tested on two challenging public datasets (ShanghaiTech, UCF_CC_50) and achieved better performance, which demonstrates the effectiveness of the proposed method.