Accurate SLAM with accurate semantic tagging would be a big deal, yes, but this project still relies on pretrained data with COLMAP so how is that relevant to your comment?
I'm not saying you can use THIS method for a SLAM; I'm saying you can use this METHOD for a slam.
You don't need a perfect COLMAP for this method (well, not this in particular, but for this method more broadly with some modifications); you just need an approximate location for a few of the images to start and then match the others progressively... which is literally what SLAM is all about.
And "pre-trained data" makes no sense. It's trained, as in slowly chewing iteratively on the data before getting decent 3D space, but that just means it's a bit slow. Hence, my mention of simple robots that move in a semi-fixed environment rather than being unusable for self-driving.
But more broadly, it's a method to describe the real-world appearance 3d space, which may have computational and flexibility advantages over massive point clouds.
