Hi Hugues,
Since the publishment of your work KPConv, it has turned out to be a classic and popular backbone in 3D processing now.
While I'm currently trying to study KPConv in more detail, I find that the processing of ShapeNetPart is a bit hard to understand. Especially, is there any specific reason to switch the y-z axis in datasets/ShapeNetPart.py#L251C1-L252C1 ?
Besides, I find the normalization of the object point cloud uses the (min-max)/2 as the centroid position, instead of the normal use of the mean position of all points.
Best，
Liyao
Hi Hugues,
Since the publishment of your work KPConv, it has turned out to be a classic and popular backbone in 3D processing now.
While I'm currently trying to study KPConv in more detail, I find that the processing of ShapeNetPart is a bit hard to understand. Especially, is there any specific reason to switch the y-z axis in
datasets/ShapeNetPart.py#L251C1-L252C1?Besides, I find the normalization of the object point cloud uses the (min-max)/2 as the centroid position, instead of the normal use of the mean position of all points.
Best，
Liyao