About the point selection

[Monad-Cube]

Thank you for your great research.
I have some confusion in the code to bother you:
Firstly, I wonder the reason for setting five TPM layers and ref_filter_steps[1,2,3,4]. What is the purpose to add a 0head without selecting points?
Additionally, when visualizing the point and bbox or computing loss, it seems that the model use 64 points to predict final bbox rather than 32 points described in paper, will 64 perform better than 32?