Crowd Counting Via Residual Building Block Convolutional Neural Network

Abstract

We present a new method called residual building block convolutional neural network (RBB-CNN) for generating high-quality density maps and count estimation by applying stacked residual building blocks. The specific deploy of convolution layers in building blocks are inspired by the work of VGG16. The RBB-CNN is an easy-trained end-to-end model and allows arbitrary-size input because of its pure convolutional structure. To verify the validation of the residual building block, an ablation on ShanghaiTech Part-A is implemented. Meanwhile, we demonstrate the performance of RBB-CNN on three crowd counting datasets, i.e., ShanghaiTech, UCSD and MALL. With a wide range from dense to sparse density, our model achieves the state-of-the-art performance on all of the above datasets.