In case the code does not run

[xqyd]

As mentioned by the author, it is better to reduce your points cloud around 95k points. Say you can use Meshlab to simplify your original ply file. After that, feed the simplified points cloud in ply format to the run.py, print the QUERY_EACH to make sure it is 10 (this will generate the npz file for downstream processing). And you can get the final result on a single 2080Ti (11GB) without memory issue. Hope this will help those who want to give this wonderful job a try. Peace!