The mysterious 'bowl effect'

[patcarbon]

I've seen a few references to this bowl effect on pages here.  i've also seen what looks like the bowl effect in data.  In fact, this was in images taken with the UAV produced by the Swedish re-seller of Photoscan, SmartPlanes.

What causes this?  Since the effect coveres a whole scene, in my case with over 200 images, I have always assumed that it can't be lens distortion since the bowl fits a 2nd order surface.  When is this distorsion introduced?  Alignment? Build Geometry? Georeferencing?  Surely it's in one of the first 2 since, if I understand right, georeferencing just uses a rigid 7-parameter transform so that means the 'bowl has to be generated earlier'.  Correct?

And what's the maths behind the optimisation?  It it relying on a photogrammetry-style epipolar constraints approach?

Patrice

[Alexey Pasumansky]

Hello Patrice,
 
The bowl effect can be introduced during photo alignment, in case camera calibration estimates are inaccurate. If camera calibration is known in advance, it can be loaded in PhotoScan and fixed during photo alignment. But the recommended approach is to correct possible bowl effect during optimization procedure based on camera or GCP coordinates, performed after photo alignment. Please note that it is recommended to perform optimization based on ground control data in any case, even if precalibrated cameras are used.

During optimization PhotoScan performs full photogrammetric adjustment taking into account additional constraints introduced by ground control data. Extrinsic and intrinsic parameters for all cameras are optimized at this step, in contrast to the simple 7-parameter transform used for georeferencing by default. Thus optimization helps to significantly improve accuracy of the final solution.

[patcarbon]

Thanks Alexey

Do you think optimisation will perform well if I use about 150 camera positions from a small, non-corrected, GPS?  Again this is from the SmartPlanes UAV.

Patrice

[Alexey Pasumansky]

Hello Patrice,

Yes, we assume that optimization will fix the "bowl-effect" even GPS coordinates are not so accurate (about several meters).

[mcstieg]

Hello Alexey!

Would you be so kind and tell me how exactly I can improve projects with bowl effect in it?

1) How should I take photos of an additional model for finding good camera parameters?

2) At which stage of my model should I import known values?

3) What do I need set as fixed and which parameters should be calculated in the model?

4) Which parameters of optimization should be activated for fisheye cameras if I want to improve the model... And if I change focal length manually?

Thank you very very much!

Edit: Just wanted to tell you that I use markers (23) in all of my models. Hopefully they can help now.