3 convolutional layers instead of 2 per level to extract feature pyramid perform better?

[limpbot]

Hi guys, in your pytorch implementation you are using 3 convolutional layers per level to extract the feature pyramid.
In the paper you only talk about two layers, are there insights that 3 layers perform better?

Best,
Leo

[limpbot]

ok i found it in the ablation studies, my bad.