UNET++ with Scale Pyramid for Crowd Counting
Crowd counting is a popular study that offers beneficial applications in various fields. Despite the benefits, developing a crowd counting application with high accuracy is challenging because of the variation in the training images, such as the density of the crowd, the perspective distortion, and the camera position. To address this challenge, a better crowd counting method needs to be developed. In this study, we propose a deep learning method based on UNet++ for crowd counting. We used VGG as the backbone and add the Scale Pyramid module as the transition between VGG and UNet++. The result of the experiment reveals that our proposed method achieved state-of-the-art performance in the ShanghaiTech Part A dataset with an MAE of 54.8 and an MSE of 85.4.
ICIC Express Letters
Marcellino, Tjeng Wawan Cenggoro, Bens Pardamean