Crowd Counting with Distribution Matching and Dilated Networks

Crowd counting is an important task due to its various application areas. In this study, we used deep convolutional neural networks to estimate the crowd count. By combining a suitable loss function and distribution matching, we propose an original architecture which increases not only counting accuracy, especially for congested scenes, but robustness across datasets as well. These improvements have been shown on Shanghaitech and UCF datasets in terms of mean absolute error and root mean square error measures. Based on the same measures, our model also exhibits superiority against many recent studies especially on the UCF dataset, which consists of highly congested scenes.