Zero-shot classification for ScanObjectNN

[kaimingkuang]

Thanks for the amazing work! I modified the zero-shot classification code for ScanObjectNN, but only got 10.5% (15.38% in the paper) accuracy for the OBJ_ONLY variation. Is there anything missing (different coordinate normalization, default point-cloud orientation, etc) in the ModelNet40 pipeline that is necessary for reproducing ScanObjectNN?