Performance discrepancy on ScanNet reproduction

[ai-cho]

Hi,

I’ve been trying to reproduce the ScanNet benchmarks reported in your paper, but I’m consistently seeing a significant performance drop compared to the reported values. I followed the provided train & evaluation scripts, yet the gap remains substantial across all class configurations.

I noticed a similar issue has been reported here: #9

Performance Comparison

Class	Reported (mIoU / mAcc)	Reproduced (mIoU / mAcc)
19 classes	32.5 / 49.1	23.15 / 38.03
15 classes	35.5 / 53.5	27.79 / 44.20
10 classes	42.6 / 63.2	33.51 / 50.56

Could you please provide some insight or solution into the following?