A preliminary benchmark of four saliency algorithms on comic art

Predicting the salient regions of a comic book panel has the potential to drive a variety of applications such as segmentation, cropping, effects such as moves on stills, etc. Computational saliency algorithms have been widely tested on a variety of natural images, and extensively benchmarked. We report the performance of four saliency algorithms on a set of comic panels taken from public domain legacy comics. We find that a data-driven method performs highest based on two metrics, Normalized Scanpath Saliency and Area Under the Curve. We discuss possible reasons for this finding based on an exploratory analysis of the similarity between the comic images in our dataset and images used in the dataset of the data driven method.