squared distances

[Ph.D. in Machine Learning and Computer Vision]

First of all thanks for sharing you RandLa-Net implementation, it is very helpful.
I've noticed that the distances returned by knn() are squared (at least when using torch_points_kernels).
In the paper they mention using non-squared distances, and from a quick look at the tensorflow code, it seems they indeed use non-squared distances.
Hope this is helpful.
Thanks,
Amnon Drory

[Tibo]

Thanks Amnon for your comment !

Indeed, this could be an important difference that impoverished our results.
Did you noticed any improvement in results with the non-squared distance ?

Thibaud

[eyalol]

@AmnonDrory , Did you manage to try using un-squared distances? was there an improvement?
I've noticed that the torch_point_kernels knn returns automatically squared distances, so I'm wondering if there is any other way fixing this issue rather than just taking the square root of distances.
Thanks