Dear authors,
thanks for releasing this amazing code base!
NYU-v2 (single-frame depth)
I am trying to reproduce the single-frame depth performance on NYU-v2 reported in Table 3:
Here are the results I get by following the evaluation instruction here. As you can see, there is a consistent gap from the reported numbers for both methods.
I am wondering if you have any idea of what could be wrong.
ScanNet (Camera Pose)
I also tried to reproduce performance on 100 ScanNet scenes following the instruction here.
With the default hyperparameters, DUSt3R and MonSt3R failed on a different number of scenes:
I am wondering how you handled this to get the numbers in the paper. Did you adjust the flow loss weight for each scene to avoid failure or only compute metrics on common scenes that all methods succeed?
Dear authors,
thanks for releasing this amazing code base!
NYU-v2 (single-frame depth)
I am trying to reproduce the single-frame depth performance on NYU-v2 reported in Table 3:
Here are the results I get by following the evaluation instruction here. As you can see, there is a consistent gap from the reported numbers for both methods.
I am wondering if you have any idea of what could be wrong.
ScanNet (Camera Pose)
I also tried to reproduce performance on 100 ScanNet scenes following the instruction here.
With the default hyperparameters, DUSt3R and MonSt3R failed on a different number of scenes:
I am wondering how you handled this to get the numbers in the paper. Did you adjust the flow loss weight for each scene to avoid failure or only compute metrics on common scenes that all methods succeed?