> Is the spatial memory a fixed size (how big?) or does it grow over time?

Fixed number of parameters

> And is there a point at which is is saturated and future results decline?

The model is statistical, so precision will steady improve, asymptotically approaching a maximum

> Two authors (so, one, lol)!

What do you mean?

There would not be much point. Colmap is already very capable in reconstructing a 3D scene from images from unknown poses if you have the camera intrinsics.

Besides processing speed, this project (and the underlying dust3r model) strength is that it works with very few images. You basically just need 2, and it can infer pseudo instrinsics and matching extrinsics on it's own.

I don't see why it could not be adapted to output gaussian splats instead. As a matter of fact, it's already been done with dust3r: https://github.com/nerlfield/wild-gaussian-splatting.

This relies on Dust3r underneath as part of its stack (I didn’t read carefully enough to tell you if it’s training or inference but I think it’s training), which outputs splats. What’s special about this is that it outputs really dense nice point clouds with arbitrary photos. We have a lot more tools that work well with point clouds than with splats, so this is nice work.