OpenSFM slow for large linear datasets

OpenSFM starts to take a long time for each additional image for me for longer sequences. Using a sequence of 700 images, the bundle stage starts to take closer to 20 seconds per image, rather than more like 2 seconds per image at the start.

The data I'm using is video, so I know that my images are all in order and only e.g. need to be matched with nearby frames. I don't have exif data, so something like matching_time_neighbors isn't applicable, but I've implemented my own similar config setting that should only take the 3 frames before and after. This indeed speeds up the matching substantially (since it only looks for matching pairs in nearby frames), but the bundle time per frame still grows.

Should it still grow? I presume the bundle is adjusting all of the points in our construction so far, but this seems like possible overkill, since my new points should only be added on the end.

Advice on config settings/directions to look at code to change for linear sequences would be appreciated!

Start by tracing the bundle stage and the nearby-frame matching configuration described in the issue. Compare how bundle time scales as a sequence grows, then determine what a linear-sequence configuration or approach would need to change; done should include validated scaling on a large, ordered image sequence.