Questions Regarding GT Scale Generation and Pose Origin in ScanNet-GSReg Dataset

[Lee-JaeWon]

Thank you very much for providing the ScanNet-GSReg dataset.

I would like to ask a few questions regarding the ground-truth scale:

Could you please explain how the GT scale was generated for each scene pair?
Specifically, I am interested in the process or methodology you used to determine the ground-truth scale between partially overlapping scene sets.

Additionally, are the camera poses for each scene based on COLMAP reconstruction, or was a different method used for pose estimation?

Your clarification would be greatly appreciated. Thank you in advance for your support.

[Jiahao620]

The relative GT scale is randomly sampled from 1/5 to 5.
The camera poses I adopted are the GT camera poses from the original ScanNet dataset.

[Lee-JaeWon]

@Jiahao620 Is it A/B? Or is it B/A?

[Lee-JaeWon]

@Jiahao620 And I would appreciate your help with #23 as well. I suspect there may be some issues, because even when strictly following the same instructions, the results cannot be reproduced, and it seems that others are experiencing the same problem.

[Lee-JaeWon]

@Jiahao620
If the Gaussian primitives are not generated from RGB-D data, I believe that using the scale factor applied to the GT pose as the ground truth for the registration scale is problematic. In my experiments, applying the gt_scale provided in the test_transformations.npz(gt_transformations_list) file to the source point cloud did not result in correct alignment.
Even if the GT pose is scaled and used during the optimization of 3D Gaussian Splatting, I believe the xyz coordinates of the Gaussian primitives remain arbitrary, and thus do not guarantee a consistent metric scale.
I would appreciate it if you could confirm whether my understanding is correct. I aim to evaluate the RSE fairly and to cite your excellent paper accurately.
For reference, all of my tests have been conducted on the 82 test sets from ScanNet-GSReg.

[makabakayyy]

Thank you very much for providing the ScanNet-GSReg dataset.

I would like to ask a few questions regarding the ground-truth scale:

Could you please explain how the GT scale was generated for each scene pair? Specifically, I am interested in the process or methodology you used to determine the ground-truth scale between partially overlapping scene sets.

Additionally, are the camera poses for each scene based on COLMAP reconstruction, or was a different method used for pose estimation?

Your clarification would be greatly appreciated. Thank you in advance for your support.
How should I download the ScanNet-GSReg dataset? I always make mistakes when downloading