Support for different panorama image size

[Golbstein]

Hi, it seems like the inference works only with 320x640 images. Do you have plans to release a general model or have you seen that this image size is sufficient for most scenarios?

[jdk9405]

Using different image resolutions may lead to suboptimal performance in feature matching.
We used a width of 640 during training due to computational constraints, and found it sufficient in our experiments on the Matterport3D and Stanford2D3D datasets.

I plan to release a reimplementation of the EDM training code in the future, though currently I’m not sure when that will be possible.