Hi, thank you for releasing the code!
I am trying to reproduce the Tanks and Temples benchmark results for the RaDe-GS + Gaussian Wrapping setting. I ran the TNT benchmark on the six default scenes (Barn, Caterpillar, Courthouse, Ignatius, Meetingroom, Truck) using the provided benchmark_tnt_gw_radegs.py pipeline, with -r 2, --data_device cpu, and without --depth_order.
The full-scene Ours (2p) results match the paper very closely, so the dataset and evaluation setup seem mostly correct. However, for RaDe-GS + Gaussian Wrapping, my virtual scan mean F-score is lower than the paper/Table 3:
| Method |
Eval |
Precision |
Recall |
F-score |
| radegs + GW |
virtual_scan |
0.4159 |
0.4588 |
0.4313 |
The paper reports around 0.48 mean virtual scan F-score for RaDe-GS + Gaussian Wrapping.
Could you clarify whether the released benchmark_tnt_gw_radegs.py exactly matches the configuration used for Table 3? In particular, are there any additional settings for mesh extraction, isosurface_value, post-processing, random seeds, data preprocessing, or RaDe-GS training parameters that are needed to reproduce the reported score?
Thanks!
Hi, thank you for releasing the code!
I am trying to reproduce the Tanks and Temples benchmark results for the RaDe-GS + Gaussian Wrapping setting. I ran the TNT benchmark on the six default scenes (
Barn,Caterpillar,Courthouse,Ignatius,Meetingroom,Truck) using the providedbenchmark_tnt_gw_radegs.pypipeline, with-r 2,--data_device cpu, and without--depth_order.The full-scene
Ours (2p)results match the paper very closely, so the dataset and evaluation setup seem mostly correct. However, forRaDe-GS + Gaussian Wrapping, my virtual scan mean F-score is lower than the paper/Table 3:The paper reports around
0.48mean virtual scan F-score forRaDe-GS + Gaussian Wrapping.Could you clarify whether the released
benchmark_tnt_gw_radegs.pyexactly matches the configuration used for Table 3? In particular, are there any additional settings for mesh extraction,isosurface_value, post-processing, random seeds, data preprocessing, or RaDe-GS training parameters that are needed to reproduce the reported score?Thanks!