Why use the relative position encoding
yifliu3 opened this issue · 1 comments
yifliu3 commented
Hi,
I noticed that you have used the relative position of centroids as the input of the position encoding:
entroid_delta = centroids.unsqueeze(-1) - centroids.unsqueeze(-2)
may I know the reason to use relative encoding rather the commonly used absolution position encoding
eliphatfs commented
I think this is common since PointNet++. This makes the patches independent of the absolute position so the network can embed it more generalizably.