A Improved Generative Adversarial Networking Architectures for Crowd Counting

Crowd counting on computer vision targets at the statistics for the number of people within an image. At present, the problem generally solved by estimating the amount of probability annotation dots within density map generated from the crowd scene. We propose a IGAN(Improved Generative Adversarial Networking) architecture to solve the accuracy of generative density map. The Improved GAN included two parts: generater and discriminator. In the generater, we predict the density maps by inputting images to the generative network. And the work of discriminator is to distinguish features between the generative density map and corresponding given veritable map, and force the generater to produce rational density map as close as the ground truth. The Improved GAN is trained collaboratively by blending together density loss and adversarial loss. Through experiments on five popular public datasets(ShanghaiTech PartA and PartB, WorldEXPO'10, UCF_CC_50 and UCSD), we validate the preferable performance of the Improved GAN in the complex scene.