|Description (include details on usage, files and paper references)||The Automatic Image Cropping dataset contains ill-composed images with manual crops provided by qualified experts.
As described in Section 2.1, our visual composition model is trained on carefully selected, well-composed photographs. Specifically, we download images from Photo.net, remove low-quality images, and collected 3000 high-quality, good composition photos for training. On the other hand, we collect 500 ill-composed photographs, which are then cropped by 10 expert users on Amazon Mechanical Turk who passed a strict qualification test. We call this labelled dataset as human crop dataset.
We use all the 3000 images to train both our system and
. To train the models of  which require both before and
after-crop images, the training split of human crop dataset
is used. All the evaluation are carried out on the test split
of the human crop dataset.
Automatic Image Cropping using Visual Composition, Boundary Simplicity and Content Preservation Models
Chen Fang, Zhe Lin, Radomir Mech, and Xiaohui Shen
ACM Multimedia (MM) 2014
Algorithm developed in this paper has been chosen to be shipped with couple of products in Photoshop, enabling new automatic image cropping feature! US patents pending.
[PDF] [Supplemental] [Project Page] [Dataset:crops] [Dataset:images] [BiBTex]