Performance Recall@1 COCO (table 6)

[FiorenzoParascandolo1]

I'm trying to calculate the recall@1 on COCO of the paper in the title (table 6).
I have some questions since I'm not getting the same results shown in the paper:

• Are the results calculated on the Karpathy split test?
• An image, caption pair is considered a true positive if 1 of the 5 captions associated with such image is in the top-1?
  I tried using the Karpathy split test. I calculated the recall@1 but I get the following results:

CLIP (openai): 52 vs 50 in the paper
NegCLIP: 58 vs 56 in the paper