Data and Code for Paper "Expressing Visual Relationships via Language"
Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mohit Bansal
The slide of our talk in ACL 2019 is also available here.
Image Editing Corpus Dataset
Json files in folder
data/ are three splits of the
Descriptions of each field:
img0: The name of image to be edited. img1: The name of image after editing. sents: A list of image editing requests. One for training and three for valid / test. uid: Universal id for each instruction. It is useful when multiple datasets are involved.
Image pairs are now available on google drive. link to download.
- Script to convert/unify NLVR2 and Spot-the-Diff datasets.
- Reproducible code in PyTorch.
Hao Tan: firstname.lastname@example.org
Franck Dernoncourt: email@example.com
We specially thank Nham Le in collecting the dataset!