Unable to reproduce SAM3D / MV-SAM3D results on GSO30 — benchmarking protocol?

[s-esposito]

Hi, thanks for releasing this work!
I've been trying to reproduce the SAM3D and MV-SAM3D numbers reported on GSO30, but my results don't match. Before assuming the gap is on my end, could you share more details about the evaluation setup and how to reproduce it? An evaluation script would be ideal, but any documentation of the exact protocol would help a lot.
Happy to share my current setup and numbers if useful. Thanks!

[wuhoward]

I also observed quality degradation on "clean" images with a clear central object and white background. I tried plotting the 3d weights, but the distribution doesn't quite behave like the ones shown in the paper somehow.

[s-esposito]

Would be great if the authors could comment :)