Autocropping: A closer look at benchmark datasets

Image cropping is a common image editing task that aims to improve the composition as well as the aesthetics of an image by extracting well-composed sub-regions of the original image. For choosing the “best” autocropping method it is therefore important to consider on which datasets this method is validated and possibly trained. In this work we conduct a detailed analysis of the main datasets in the state of the art in terms of statistics, diversity and coverage of the selected sub-regions, namely the ground-truth candidate views. An analysis of how much semantics of ground-truth candidate views is preserved with respect to original images and a comparison among dummy autocropping solutions and state of the art methods is also presented and discussed. Results show that each dataset models the cropping problem differently, and in some cases very high performance can be reached by using a dummy autocropping strategy.