Control points acting erratically, illustrated

Below you can see I’ve added a control point across 14 images.  The photos are in sharp focus and the points are reasonably unambiguous, so I saw an error of less than a pixel while placing them.

After aligning, they all show wildly high error – e.g. ‘IMG_8965_DxO.jpg’ has an error of 137 billion (!) pixels, even though we can see the point is placed correctly (see the purple coded view below):

After deleting that point, the rest of the CPs snapped back to ~1 pixel error.

I then added the offending point back, where it now shows as 0.2 pixel error as below:

If I move any of these points even slightly, all of their error values move to exactly 100 pixels, instantly.

To make this even more mysterious, this isn’t a complex or corrupt file; I started with a clean cache and new project – I simply:

  • Imported photos
  • Imported points
  • Launched alignment

Has anyone seen this before?

No, I have not.I have seen ridiculously high errors and also the notorious 100 in a VERY difficult project (images pulled from the net) but they always went away after alignment. I also did not notice the correlation to just one point.

I would definitely call that a bug!

Do you still have the state with the offending CP? If yes, how do the distortion grids look like? Does it go away if you delete the weird component and align it again? The thing is that there is some randomness in the alignment process which gets noticable only when the model is borderline in terms of finding a solution (not quite ideal) - which is not surprising given the amount of metal and unicolor surfaces…

Hello, Götz,

As RC has been crashing regularly the past days during alignment on this project, I don’t have the intermediate state shown above but will take a look at the distortion grids as you suggest if I can get another clean example.

I’m trying to understand why RC keeps crashing – clearly it relates to control points since I am starting with the same image set in a fresh project each time:  the project used to align fine, but after adding needed control points to address alignment deficiencies, I haven’t been able to run a single align without one (or several) core dumps.  The control points are all ‘legal’, and I’ve verified them several times.

I suspected that this could be factor – but it seems from your comment that alignment cured your obscenely high pixel error, and did not crash the program.

Yes, definitely no crashes, at least no core dumps, maybe an unexpected propram termination or so, but that can always happen.

So you didn’t have any issues with other projects before? Can you still run/align those without issues? That would exclude some problem with your setup.

Yes, it’s not the system.  I can go back to previous versions of this project, and it aligns fine.

But once I have ~200 total CPs, RC either fails to complete (stalling after several hours) or crashes outright.

After trying for several days, it’s getting very depressing.