About the use of pre-trained models

[ybc-ybc]

Hi zhu, you did a really great job!

I would like to ask a question about the use of pre-trained models. Previously, due to the lack of pre-trained models in the point cloud domain, different approaches used different backbone networks and preprocessing in weakly supervised point cloud learning. This resulted in different methods being hard to align and very confusing. I wonder if it is possible to directly use your pre-training model (Only 3D encoder and decoder ) as a base and then train the downstream datasets (S3DIS and ScanNet) under weak supervision.

[HaoyiZhu]

Thanks for your interest. I think it is absolutely okay. You just need to load the pretrained backbone as initialization.