the meaning of neighborhood_limits

[cppcute-pm]

Hello, HuguesTHOMAS.
Thanks for your amazing work! In every calibration function of all the sampler classes, you remarked that：
Neighbors calibration: Set the "neighborhood_limits" (the maximum number of neighbors allowed in convolutions)
so that 90% of the neighborhoods remain untouched. There is a limit for each layer.
I'dont understand why some parts of the neighborhood points should remain untouched. And I want to use the KPConv module in my experiment and try to build a simpler dataloader, I'm wandering that whether it necessary to add a calibration function to gain the batch_limit and neighborhood_limits or I could just set the batch_num and neighborhood_num fixed with fixing number of downsampled points.

[HuguesTHOMAS]

You definitely can. The calibration function is here to simplify the parameter tuning by choosing an appropriate value depending on the point density. But you can set it yourself absolutely.

[cppcute-pm]

Thanks for your reply! One more thing, could you explain why the calibration function needs the 90% of the neighborhoods remain untouched ? I don't understand.

[HuguesTHOMAS]

It does not "need" it. It is just a value you can decide for yourself. In my experiments, I noticed this was a good value as it already reduced the computations by a lot, while not losing any performance

[cppcute-pm]

Thanks for your reply！