Soccer Video Event Detection Using Metric Learning

Authors

Ali Karimi, Ramin Toosi, Mohammad Ali Akhaee

Abstract

We presented a novel video-based event detection method in soccer videos. Our proposed method is based on convolutional neural networks(CNNs) and gated recurrent units(GRUs) networks and uses contrastive loss for network training. We also collected a dataset called SoccerEvent to perform various experiments. To achieve the best choice of architecture for CNN and recurrent neural networks(RNNs), Various experiments were performed on different architectures on the SoccerEvent dataset. Then we selected the best architecture for the proposed method. Experiments conducted on two different datasets, SoccerNet and SoccerEvent, show that the proposed method has achieved higher accuracy than other methods. The proposed method has reached 64.9% accuracy in the SoccerNet dataset and 93.76% in the SoccerEvent dataset, which is about 1% better than other state-of-the-art methods.