Point cloud without normals

[rvxfahim]

Hi, I tried to do segmentation on raw point clouds without any normals, but it failed, saying there are no normals available.
Is it a strict requirement to provide the normals also ? Or is it possible to do segmentation without normals somehow?

[Gofinge]

Usually, it won't be too difficult to extract normal (view-direction) from the original data format, e.g., even for point clouds lifted from Video by VGGT, we have a good pipeline to extract normal, which we will release as a demo in follow-up research of Sonata (Concerto).

However, just in case we cannot access color/normal, whatever, in the next research, we will add RandomDropColor and RandomDropNormal during training, to make the model more robust in case this information is absent. At that time, just pad normal with 0 might helpful.

[Gofinge]

Hey guys, sorry for the long wait. Here are some updates:

Since Concerto, we already support w/o color and w/o normal by training the model with RandomDropColor and RandomDropNormal. Simply pad color and normal with [0, 0, 0] will also get a very good performance. Note that this is a pre-release feature of Project Utonia (the next project after Concerto, all point clouds for one model). Here are some useful links:

1. Demo for support without color/normal: https://github.com/Pointcept/Concerto/blob/main/demo/0_pca.py#L44
2. Concerto weight (support without color/normal): https://huggingface.co/Pointcept/Concerto/tree/main
3. Updated Sonata weight (support without color/normal): https://huggingface.co/Pointcept/Sonata/blob/main/sonata_20251024.pth