Difficulty reproducing results with SemanticKTTI

[jimmys-antoman]

I am trying to reproduce the results of the baseline model depicted in SDSeg3D with the SemanticKITTI dataset. I am using the default settings.

However, I am getting very low mIoU ranging between 16-23%.

I am training the model in one GPU GeForce RTX 3090, for 10 epochs (although I've tried running it for 9, 13, 17 epochs) with 2 batches. I've also tried distributed training with 2 GPUs for 10 epochs with 3 batches each but it did not show any substantial improvement.

Could please provide some guidance for improving the mIoU to reach that of the baseline model?

Thank you for your time and assistance.