Recall (R@k) way lower than the one obtained in papers

[asczdvefb42]

Here is what i do :

• Get test dataset of image - caption (IIITD-20K dataset)
• calculate embeddings with my fine tuned CLIP
• calculate cosine distance between each text to all images
• get the k closest images to the text, if the corresponding image is in it, do +1 to score
• get the recall by dividing the score by the length of the test dataset.

This is my recall at k. I obtain a R@1 of 17%, while most papers when finetuning CLIP obtain at least 60% recall at 1. Any idea what i could be doing wrong?