Localized Narratives Benchmarking

Hi @JiahuiYu / @jasonbaldridge, I was wondering what "oversampling" in this line from the paper meant: "The validation set of the Localized Narratives COCO split contains only 5,000 unique images, so we follow [47] in oversampling the captions to acquire 30,000 generated images.”

Do you sample 6 images for each unique caption to generate 30k images, or do you generate 1 image for some modification of each unique caption (eg take n random sentences)?

Start with the quoted Localized Narratives COCO validation-set statement in the issue and consult the referenced paper [47] or benchmark documentation. Done means clarifying what “oversampling” means and how the 30,000 generated images are obtained.

Written by the indexing model from the issue text.