Lightweight convolutional neural networks for player detection and classification

Vision-based player detection and classification are important in sports applications. Accuracy, efficiency, and low memory consumption are desirable for real-time tasks such as intelligent broadcasts and event classification. In this paper, we present a convolutional neural network (CNN) that satisfies all these requirements. The network contains a three-branch proposal network and a four-cascade classification network. Our method first trains these cascaded networks from labeled image patches. Then, we efficiently apply the network to a whole image by using a dilation strategy in testing. We conducted experiments on soccer, basketball, ice hockey and pedestrian datasets. Experimental results demonstrate that our method can accurately detect players under challenging conditions. Compared with CNNs that are adapted from general object detection networks such as Faster-RCNN, our approach achieves state-of-the-art accuracy on three types of games (basketball, soccer and ice hockey) with 1000 x fewer parameters. The generality of our method is also demonstrated on a standard pedestrian detection dataset in which our method achieves competitive performance compared with state-of-the-art methods.