Performance-enhancing network pruning for crowd counting

The Counting Convolutional Neural Network (CCNN) has been widely used for crowd counting. However, they typically end up with a complicated network model resulting in a challenge for real-time processing. Existing solutions aim to reduce the size of the network model, but unavoidably sacrifice the network accuracy. Different from existing pruning solutions, in this paper, a new pruning strategy is proposed by considering the contributions of various filters to the final result. The filters in the original CCNN model are grouped into positive, negative and irrelevant types. We prune the irrelevant filters of which feature maps contain little information, and the negative filters determined by a mask learned from the training dataset. Our solution improves the results of the counting model without fine-tuning or retraining the pruned model. We demonstrate the advantages of our proposed approach on the problem of crowd counting. Our experimental results on benchmark datasets show that the network model pruned using our approach not only reduces the network size but also improves the counting accuracies by 4% to 17% less MAE than the state of the arts.