Intrinsics estimation on ultrawide FOV

Hi,

Excellent work, thanks for the detailed paper and prompt model release!

I think that the method systematically underestimates focal length on ultrawide images. For example, here is an iPhone 13 Pro ultrawide image:

MoGe estimates a focal length of 1767.9873 pixels, or about 97 degrees hFOV. But the GT hFOV for this device is about 108 degrees. In testing I have done, there seems to be a pretty consistent bias to underestimate on these.

I don't know of a nice public dataset of images of this kind to show this to you. But I can produce a related behavior by taking a public dataset with known intrinsics (Cambridge Landmarks Dataset) and cropping in the vertical. As the crop gets more extreme, MoGe's estimate gets biased towards being big:

(I attached a jupyter notebook you can use to reproduce these plots: cambridge intrinsics.ipynb.zip).

BTW, it's not reproduced here, but taking a cambridge image and cropping the per-pixel pointcloud, rather than the image itself, does not reproduce this behavior. So, I think it is not a problem with the LM solver step.

I wonder if you have any comment or ideas for improving the estimation accuracy here? Is it just that these weird crops/FOVs are well outside the training distribution?

Start with the attached cambridge intrinsics.ipynb notebook and the Cambridge Landmarks Dataset to reproduce the focal-length bias under increasingly extreme vertical crops. Compare image crops with the reported point-cloud crop behavior, then determine whether the ultrawide estimation issue can be addressed and validate any proposed improvement on these cases.