Questions about point cloud numbers

[Xin-97]

Hi, thanks for the great work!
I've encountered some issues when running the demos. It seems that the number of GPUs should be a factor of the number of objects to avoid bugs in the distributed processing setup.
I also have a question about input point numbers: How does the number of input points influence the performance? In your paper, you mentioned using 100,000 points per shape. Would the performance degrade with fewer points?

[Colin97]

We train the model by sampling 100,000 points and have not tested it with fewer. However, 100,000 points should not be a bottleneck of inference.