Coarse-to-Fine Network for Crowd Counting
Zhiyuan Sun · 2022 IEEE International Conference on Electrical Engineering, Big Data and Algorithms (EEBDA) · 2022
Crowd counting task aims at predicting the count of people in images or videos. Although recent methods perform well by using CNN, most methods only use the information contained in deep layers to predict the final number. To utilize the rich information existed in layers of different level, we propose a simple framework, i.e., Coarse-to-Fine Network (C2FNet), which stacks a set of blocks. In our method, we predict the density maps not only by the final extracted features, but using the features extracted by middle blocks as well. Therefore, we adopt a coarse-to-fine loss (C2F loss) to supervise the learning progress of coarse density maps so that our network can predict the final density map from coarse to fine. Moreover, there is a multi-scale perception module (MSPM) in each block, which extracts multi-scale features by using convolutional layers with kernels of multiple sizes and different dilation rates. Experimental results demonstrate that C2FNet has achieved great performances on ShanghaiTech Part A, ShanghaiTech Part B, UCF-QNRF and UCF-CC-50 datasets.