Hi,
During my preprocessing, I noticed that some of the images contain zero or one candidate region (i.e., image['candidate_regions']).
Also, in some cases, the referred object bounding box (i.e., annotation['bounding_box']) does not overlap any of the candidate regions, in other words, no candidate region covers the referred object.
Regarding, in many cases, the target bounding box is only a bit smaller or bigger the corresponding candidate region, adding the target bounding box as possible candidate region would make the model confused in selecting the best bounding box of the referred object.
Could you please let me know how you dealt with the above cases?

No module named pycocotools.coco

I was following the readme but I can't download the Google RefExp and UNC RefExp splits with MCG candidates, somehow the link doesn't seem to work. Could you look into it please?
Broken url: https://obj.umiacs.umd.edu/referring-expressions/google_refexp_umd_split.tar.gz

Sorry to disturb you.
The numbers in the paper are:
the training partition contains 23199 images with 67996 objects and 85,408referring expressions.
The validation partition contains 2600 images with 7623 objects and 9,602 referring expressions.

But the numbers we get from the released split is:

the training partition contains 23199 images with 44797 objects and 85,408referring expressions.
The validation partition contains 2600 images with 5023 objects and 9,602 referring expressions.
So, which one is right?