NME2D Better Than Reported in Paper

[WZUchen]

Issue Description

I've been working on reproducing the results from the DSFnet paper, and I've noticed a discrepancy in the NME2D metric. My experimental results show a lower NME2D value compared to what's reported in the paper. I'm seeking clarification on this difference.

Experimental Results

Here are the metrics I've obtained:

Metric	My Result	Paper Result
NME3D (Dense Alignment)	3.7879	3.8
NME2D	2.7141	3.27
KPT2D	2.8772	null
KPT3D	4.1352	null
REC (Reconstruction)	0.0324	0.0324
Yaw	2.65	2.65
Pitch	4.28	4.28
Roll	2.82	2.82
MAE	3.25	3.25

As you can see, all metrics match the paper's results except for NME2D.

Questions

1. Is there a specific reason why the NME2D in my experiments (2.7141) is lower than the reported value in the paper (3.27)?
2. Are there any particular preprocessing steps or evaluation protocols for NME2D that might account for this difference?
3. Could this discrepancy be due to differences in the test set or evaluation methodology?

Any insights or clarifications would be greatly appreciated. Thank you for your time and assistance in resolving this issue.