QingyongHu/SensatUrban

RandLaNet results not close to what is claimed using provided parameters

mpautzke opened this issue · 2 comments

mpautzke commented

The results I got after training RandLaNet were nowhere near the results claimed (using the parameters provided). Looking at the competition board, I see others are able to achieve similar results to the claims. Anyone able to provide a parameter set that could put me in that ballpark?

codename1995 commented

I have the same issue. I used the code of this repo but only got a max mIoU of 48.02, which is far from the reported result (58.64 in Table 3).

@QingyongHu Do you use some other settings, such as pre-trained weight, or hyperparameters different from those in tool.py?

  • 👍1
Serendipity-luck commented

I have the same issue. I used the code of this repo but only got a max mIoU of 48.02, which is far from the reported result (58.64 in Table 3).

@QingyongHu Do you use some other settings, such as pre-trained weight, or hyperparameters different from those in tool.py?

I encountered the same issue, how did you solve it?