Motion tracking: Are two keyframes or a range of frames used to calculate relative 3D positions of tracks?

Question

I'm shooting a scene where the camera will be on a tripod for the object-tracking part that'll be used in the final footage, but I'm moving the camera around first in order to allow Blender's camera tracking to calculate where the floor and wall tracking dots are.

Tracking the move-the-camera around part is tedious, and I won't be using the footage when I'm done. Which will give me the better result, or will there be no difference?

1. Track the full frame range, or

2. pick a single frame from the move-the-camera-around part to use as Keyframe A, and throw away the rest?

Some answers here seem to suggest that Blender uses a range of frames to calculate the 3D position of the tracks, which would make option 1 the better option, like this answer:

https://blender.stackexchange.com/a/27656

However, the Blender documentation kind of makes it seem like only two keyframes will be used to solve the camera and 3D position of the tracks, and that all I need for a good solve is a pair of frames that have lots of parallax differences between them:

https://docs.blender.org/manual/en/latest/movie_clip/tracking/clip/editing/track.html#solve-solution

I'm also interested in the answer to this question for object tracking: Are only two keyframes used to solve the 3D positions of the tracks relative to each other for the object, or is a range of frames used?

I do recognize that tracking the full sequence will give me a more reliable check on whether Blender did an accurate 3D calculation. I'm interested in knowing whether Blender will use the full track (or the range of frames between keyframes) to get a more accurate 3D calculation.

Answer 1

Blender uses libmv. Libmv, in turn is based on the paper Open Source 3D Reconstruction from Video by Keir Mierle. It's their master's thesis.

Two keyframes are used to approximate the 3D position of the tracking dots. Then the remainder of the frames are used to refine the approximation; if I read the source comments correctly. This is consistent with Chapter 2 of the paper and appears to be consistent with the technical description of the pipeline.

So the answer seems to be that Blender will use the full track; but how much that matters will depend on how good the initial estimate from the first two keyframes is.