FFnet: Residual Block-Based Convolutional Neural Network for Crowd Counting

Due to the nonuniform scale variations and severe occlusion, most current state-of-the-art approaches use multicolumn CNN architectures with different receptive fields to tackle these obstacles. We design a single-column network to verify the necessity of multicolumn network, and we find that under similar number of parameters and size of receptive field, single network is able to perform as well as multicolumn network. Following that, we propose a single-column network called FFnet based on residual block. FFnet is a fully convolutional network and easy to train. We perform extensive experiments on Shanghaitech dataset and the UCF_CC_50 dataset, and the results show that our method achieves a better performance than Switch-CNN with nearly half number of parameters, and a closing performance to the state-of-the-art model CP-CNN with almost one-tenth parameters.