COCO 5K test set

Open Anastasia0411 opened this issue 4 years ago • 1 comments

Hello! In your article for I2T retrieval task you report the top-K retrieval results on the 1K and 5K COCO test sets. On the official COCO page (https://cocodataset.org/#download) there are 3 versions of this dataset: 2014, 2015 and 2017, but each of these versions contains more than 5K images (41K in 2014, 81K in 2015, 41K in 2017). Could you please tell me how did you choice 5K images from 41K or 81K test sets? Thanks a lot!

Mar 03 '21 10:03 Anastasia0411

Hi Anastasia,

I'm not sure but when I download the coco_caption dataset with (like in DOWNLOAD.md)

wget https://biglmdiag.blob.core.windows.net/oscar/datasets/$TASK_NAME.zip
unzip $TASK_NAME.zip -d $DATA_DIR

The test dataset is of size 5000. It should be the choice of 5K images you are looking for?

Oct 10 '21 08:10 jontooy