[sfmTransform] rewrite auto_from_cameras

[servantftechnicolor]

SfmTransform have multiple options to "transform" the sfm cameras poses and associated point cloud.

"Auto_from_cameras" :

• translates the sfm such that the mean of the cameras positions is (0,0,0).
• scale the sfm such that the variance of the cameras positions is 1
• Rotate the sfm such that :

1. A plane is fitted to the cameras positions
2. The scene is rotated such that the plane normal become n=(1,0,0).

Previous code was not giving the correct results for the rotation part (random) and I was not able to understand it as it had no geometric meaning to me. This code is a full rewrite of the rotation estimation part.

[simogasp]

Could you please elaborate better in the description of the PR on what was not working in the previous version and the improvement you added with this PR?

[mugulmd]

Could you please elaborate better in the description of the PR on what was not working in the previous version and the improvement you added with this PR?

Did some comparison tests on multiple scenes, the results are definitely in favor of the new method:

• the new method never performs worst than the old one (sometimes the scene is upside-down, but when it is the case then it's upside-down for both methods)
• the new method solves orientation issues on most scene where the old method failed (usually upside-down problems).

Also since the new method comes with a clear and concise paper explaining it (and proving how it works mathematically) I tend to prefer it.

changes addressed

[simogasp]

Did some comparison tests on multiple scenes, the results are definitely in favor of the new method:

• the new method never performs worst than the old one (sometimes the scene is upside-down, but when it is the case then it's upside-down for both methods)
• the new method solves orientation issues on most scene where the old method failed (usually upside-down problems).

Also since the new method comes with a clear and concise paper explaining it (and proving how it works mathematically) I tend to prefer it.

Great.
It would be nice to make a dataset of all the cases you collected (just the sfmdata) and keep it somewhere, ideally for making functional/unit tests. So that in the future we can replicate the results and possibly compare again the results if a new method/heuristic is proposed.