How does this improve on taking a video with your phone and feeding it to one of the OpenSource 3D reconstruction packages (such as e.g. MeshRoom [1])?
Meshroom definitely can't do 3D reconstruction from monocular video. But there are other algorithms that can. Just Google "3D reconstruction monocular" and you get loads of demos. E.g. https://youtu.be/TZ1eToXQwN0
Unfortunately I don't know of any that are open source and not research-project style code (i.e. weeks of work to maybe get working).
Matterport are the market leaders in this sort of thing but it's commercial.
> Meshroom definitely can't do 3D reconstruction from monocular video.
It's quite possible we're not talking about the same thing here, but for my understanding of what "3D reconstruction from monocular video" means, Meshroom most certainly can do that.
To wit, I recently shot a video of a house with my drone, which happens to have just one camera. I then proceeded to extract the frames of the video and fed them to Meshroom.
I am now in possession of a full 3D textured model of said house.
[EDIT]: now that I've watched the video you posted, I see they do the reconstruction in real time, which is cool and which Meshroom can't do. However, if you aren't in a rush, the end result is the same.
I don't think it helps meshroom much because (afaict) there's no easy way to tell meshroom the angular information that's known from the stepper controller (I couldn't find anything obvious in the docs). If that worked, you could avoid the expensive and error prone initial orientation-determining step, which is absurdly slow.
I've done similar to this article but using Hugin as a panorama constructor. It speeds up the process signnificantly (because again you don't have to search for image orientations if you "know" them).
from my experience photogrammetry suites tend to produce results that are a lot more 'delusional' than measuring with an actual sensor, especially if you're photographing reflective/metallic/transparent surfaces.
Yeah, LIDAR is sensitive too, but usually the data is missing rather than random/delusional. Missing points tend to be less labor-intensive to clean up on the results than erroneous points that totally deform whatever mesh they're a member of.
Unfortunately I don't know of any that are open source and not research-project style code (i.e. weeks of work to maybe get working).
Matterport are the market leaders in this sort of thing but it's commercial.
https://matterport.com/