ScanNet dataset weights eval

[ZGY0602]

Hello, when I evaluated using the best trained weights on the ScanNet dataset, the results differed significantly from what was printed during validation when training on ScanNet. How can I resolve this?

[danielS91]

Thanks for reaching out - can you share some more details about what you did?

The original EMSANet was not trained on ScanNet - we only used EMSANet in our follow-up works EMSAFormer and Panoptic Mapping. So did you train it on ScanNet yourself, or did you use one of our checkpoints?

[ZGY0602]

Thank you for your response. I trained EMSANet on the ScanNet dataset, but during evaluation, all the metrics in the generated weight files were abnormally low.