Performance on ScanNet200 without using pre-computed masks

[ndkhanh360]

Dear authors,

I have tried to reproduce the results on ScanNet200 by generating the 3D masks with Mask3D, but the AP I obtain is consistently below 20.0, which is much lower than the reported 24.7 AP achieved with the pre-computed masks.

I followed the instructions and ran:

python run_evaluation.py --dataset_name scannet200

I’m wondering if there might be an issue on my end. Could you please provide more detailed instructions on how to obtain the pre-computed masks? I would greatly appreciate it.

Thanks, and I look forward to your reply.