Could you please provide the coco annotations and metadata used for Open-vocabulary COCO experiment?

[timqqt]

I am having a hard time to reproduce the open-vocabulary coco experiment results in Table 9 and Table 13. The performance on base categories drops more than the paper reported.

I would appreciate if you can provide annotations including instances_train2017_seen_2.json, instances_val2017_unseen_2.json, instances_val2017_seen_2.json, instances_val2017_all_2_oriorder.json, nstances_train2017_seen_2_oriorder.json and captions_train2017_tags_allcaps.json

Cheers,