Locating images in reconstructions

[BrookRoberts]

It may be useful to have an ability to find images in a reconstruction without adding them to the reconstruction. a) You may believe the existing reconstruction is already plenty accurate b) Opensfm becomes slow with large reconstructions due to bundle adjust/loading large files

Assuming no-one else makes this for me in the next week, I will implement this in some form. I don't think there is a beautiful way of just slipping this in the code - either I will have to make adjustments to lots of functions, so that they can understand how to locate images but not add them. Or I will need to duplicate code.

Would this be a useful function of opensfm as a whole? If yes, I'll go for a more intrusive solution, if not, I'll try and do a more hacky external solution (since I'd rather not have to have my own version of opensfm). Any ideas for clean ways to do this? (general idea I think would be that I would calculate features, ideally only match with images in reconstruction, although probably cost of matching with the other images to be located may be small, then add them to a new tracks file (rather than overwriting), and then locate them in reconstruction by using bundle_single_view function).

[paulinus]

@BrookRoberts what you want is to localize a new image with respect to an existing reconstruction?
There was some interest in that in #115. See #115 (comment) for a short description of how it is typically done.

Ideally, this can be implemented as a set of new commands and without modifying much of the existing code. Probably one command to build a global index for feature matching after the reconstruction is done and another command to relocalize a new image.

[BrookRoberts]

Thanks @paulinus . I have a moderately working solution for this. Assuming you have gps the first part, matching features is not a problem simply using the existing solution I made for extending recontsructions. At the moment this is fast enough for my needs (I just make a shallow copy of an existing reconstruction, and actually add it), but then I currently do the full bundle stuff, haven't yet switched to just doing bundle_single_view. That change might require me to change a few things.

I'll probably resist pushing anything based on this until the extending recontsructions pull request is in though, since the stuff I'm working on here is building on itself and I don't want my branches to get to complicated.