Question about the recall@1 results of STUN.

[euncheolChoi]

Thank you for sharing your work!

I have one question about the experimental results in this paper.
The original Netvlad records a recall at 1 of about 86% for pitts30k. However, the recall at 1 of the network proposed in this paper is measured to be about 61%.
I think a lot of this difference can be attributed to the difference in pooling layers. In this paper, it used the gem pooling layer instead of the vlad pooling layer.

Could you share any personal experience or insight into the reason for this choice?

Thank you in advance.