BA updates - save covariances, grid size -> patch size, more config options

[akshay-krishnan]

• Add functionality to compute and save the covariances using gtsam.Marginals. This is only for debugging, not for use in the algorithm.
• Min track length pre-BA should be 3.
• Do not BA cameras that have < 20 tracks. What do we we do with them? retain the pre-BA camera poses as these could be important to merge children upstream. Should we also retain the tracks? discarding them now, could revisit later if that causes problems.
• change from grid-size to patch size for filtering tracks. grid-size is unclear, perhaps this was using recangular patches before?
• expose some more options to customize BA, eg. relative cost tolerance.

[hkhanuja]

nothing wrong here, but just confused that we are taking the covariance matrices from the cameras_to_model and then set it in optimized data which is the recovered after optimization so should have those inherently?

[akshay-krishnan]

are you suggesting moving this to __optimize_and_recover ?

[hkhanuja]

no I am just a little confused on what this if block does

[hkhanuja]

Just thinking if it should be called number of tracks per camera or number of measurements per camera? and I think if we have 2 cameras with less than 20 measurements after all pre-ba filtering, we shouldn't remove those tracks because those tracks might have measurements in other cameras (maybe helpful in merging ahead)

[akshay-krishnan]

Just thinking if it should be called number of tracks per camera or number of measurements per camera? and I think if we have 2 cameras with less than 20 measurements after all pre-ba filtering, we shouldn't remove those tracks because those tracks might have measurements in other cameras (maybe helpful in merging ahead)

we dont remove the tracks, we remove the camera (which removes the measurements during BA). tracks_per_camera or measurements_per_camera is the same since it is defined for a particular camera (tracks == measurements). while measurements is more accurate, tracks is a simpler name.