An Empirical Evaluation of Cross-scene Crowd Counting Performance

Crowd counting and density estimation are useful but also challenging tasks in many video surveillance systems, especially in cross-scene settings with dense crowds, if the target scene significantly differs from the ones used for training. Recently, Convolutional Neural Networks (CNNs) have boosted the performance of crowd counting systems, but they require massive amounts of annotated training data. As a consequence, when training data is scarce or not representative of deployment scenarios, also CNNs may suffer from over-fitting to a different extent, and may hardly generalise to images coming from different scenes. In this work we focus on real-world, challenging application scenarios when no annotated crowd images from a given target scene are available, and evaluate the cross-scene effectiveness of several regression-based state-of-the-art methods, including the most recent, CNN-based ones, through extensive cross-data set experiments. Our results show that some of the existing CNN-based approaches are capable of generalising to target scenes which differ from the ones used for training in the background or lighting conditions, whereas their effectiveness considerably degrades under different perspective and scale.