Cosine Focal Loss-based Change Detection for Video Surveillance Systems

Abstract

Change detection is one of the important tasks for video surveillance systems. A variety of learning-based approaches have been proposed, but class imbalance in training data degrades their learning efficiency. In this paper, we propose a cross entropy loss with a modulating term in cosine form to handle this class imbalance. Although the original focal loss focuses only on reducing weights for well-classified data, the proposed function is designed to preserve sufficient gradients for rare hard samples as well. This property allows a network to learn mainly from a few significant samples on which the network should focus. We validate the proposed loss through various experiments on CDNet2014 dataset, and the results show that the network trained with the proposed loss achieves better performance than other state-of-the-arts in various complex scenarios.