Data for the paper "Learning to generalize to new compositions in image understanding", Yuval Atzmon, Jonathan Berant, Vahid Kezami, Amir Globerson and Gal Chechik
The format of the *split* text files is Triplet ID|Image ID|S, R, O
Triplet ID can be ignored.
Dataset was derived from annotation of the COCO dataset, and therefore we release it under the same license, which is Creative Commons Attribution 4.0 License.