About the per-point feature-aware weight

[hphnngcquan]

Hi @chenhonghua ,
It's me again, i have closed the previous issue.

For this issue i want to ask about the per-point feature-aware weight gj.
As I understand, generating gj requires a mesh model. But i plan to use RePCD-Net on KITTI dataset, which only provides x, y and z coordinates.
Have you trained the network on similar dataset as KITTI, and how do you acquire gj?

Thank you

[chenhonghua]

Hi, g_j is kind of like a geometric feature (like curvature). You can also try to compute such like feature prior based on point clouds.
Regarding KITTI, I have not tried it.