Could I use this to improve results from Colmap? Mac & Nikon camera = camera poses are all wrong.

[jared-krauss]

I'm trying to do this all locally, making a GS that I can also print 3D.

One problem I'm running into right now is that I'm capturing chaotic city night scenes. And so my colmap output is not great. And I don't know how to improve it by manually saying, no this is at 9 o clock looking center, and then no this is in between one o clock and center lookin gat 6 o clock, for instance.

So trying to find something that lets me take the current .ply output and then move the camera poses to get better .ply that I can then render using OpenSplat for the web, and which I can import to something like Rhino to get a 3d printable file. I don't know if I could do that in Blender also.

I come from analog photogrpahy. Just trying to figure out how to use 3d printing and gaussian splatting to expand the field of documentary photography.

I'm on a Mac M1 Pro 16gb :?

Edit: And this is my google doc documenting my progress so far: https://docs.google.com/document/d/1Qfyar0fL89X2R9JqByIpli1LtgETBIXKnbNzIyfjHx4/edit?usp=sharing

[SBCV]

I do not think, that there is a straight forward way using this addon to achieve what you want.
A better approach would probably be to use some kind of prior during the structure from motion process. For example, I think you could pre-define which image pairs have an actual overlap. This way you would avoid incorrect feature matches during the reconstruction process and will probably improve the resulting image registration / reconstruction

[jared-krauss]

I was wondering how to do exactly that, how to tell it which pairs have an actual overlap. I wrote out a text file to do just that, but I don't think it really helped.

I mean, tbf, my dataset is like cream of the bad crop, if you know what I mean. But the Gregg's & Nando's signs look real good when my professor renders it on his PC so somethings are overlapping correctly.

I was also thinking about creating partial splats by rough similar area or similar viewpoints, and then combine them, but wasn't sure how to do that either.

[SBCV]

I think Colmap already supports this. Read the part about custom matching.