Inquiry Regarding Inference on Unlabeled Data Using RandLA-Net

[ome-13]

I've been working with RandLA-Net, specifically mimicking the setup used for the S3DIS dataset, for semantic segmentation of point cloud data, focusing on building facades. While achieving promising results on labeled data with an m_IoU of 82%, I'm encountering challenges with inference on unlabeled data. The results aren't meeting expectations, and I believe there's room for improvement in the inference process. Could you provide insights or suggestions to enhance RandLA-Net's performance in such scenarios? Any guidance or recommended resources would be highly appreciated.

Thank you for your time and consideration.