Can you explain why the results of PCAN are lower than PN_VLAD on U.S. and R.A. in Table 3.

[sagittahjz]

Can you explain why the results of PCAN are lower than PN_VLAD on U.S. and R.A. in Table 3.

I also find the results of PCAN are lower than PN_VLAD in my 3D face recognition task.

Other components are identical except an additional attention network.

[XLechter]

Hi, @sagittahjz Sorry for later replying. I have not noticed this issue before. The reason I think is due to the datasets you used. If you look into the datasets, you will find Oxford is a complex dataset which contains a lot of noises. But U.S. and R.A. are with simple structures mainly containing trees and walls. So I think the attention moudule will work significantly on complex datasets, and will not gain more improvements on simple dataset. Hope this helps.