RoMa in Mismatched Paper

[scott-vsi]

Sorry if this is the wrong place to ask this, but I was looking at your paper, Mismatched: Evaluating the Limits of Image
Matching Approaches and Benchmarks link. In Figure 1, you put RoMa at 0% on the Niantic benchmark, but looking at Table 1 it looks like RoMa should be around 20% wrt GIM (0.104/0.560=0.18). Am I reading that right?

Also, do you have any thoughts on why RoMa did so poorly in that benchmark considering it did so well in the OoD benchmark? Is the thought that RoMa may do better if something like colmap instead of mickey were use to estimate the camera?

[scott-vsi]

I see. RoMa is at 0% because it is normalized between min/max not just max. Still, RoMa's poor performance on Niantic is surprising.