Optimization generates heavy noise in sparse cloud

[nadar]

I have 12 vertical aerial photos of a castle (acquired with a 28mm lense from about 800m flying heigth) and 150 oblique views acquired with a 135mm from different angles (2 x 360°) at flying heigth varying between 300 and 150m. All photos are well exposed and sharp. They are all referenced (GPS data stored in Exif)

I import all photos in the same chunk. They all align fine (using "ground control" or "disabled" option).
I have a few (4) ground control points, not very accurate (obtained from Google Earth). Alignment error is about 5m. I add 20 markers using guided approach, and on most photos, proposed placement is fairly good, and when needed, I have adjusted the placement manually.
The resulting sparse cloud is not too bad (noisy, but geometrically coherent).

If I optimize this sparse cloud, either using fixed camera calibration or not, the "optimized" cloud becomes very noisy, with many pikes and wrongly oriented sub-parts. As you may expect, derived dense cloud is not good...

Any suggestion will be welcome.

[Alexey Pasumansky]

Hello nadar,

Probably, these not very accurate GCPs cause the problems, as their positions are using in the optimization process with the weight specified in the Ground Control pane setting dialog.

I can suggest to export markers to xml file, then remove all of them and then optimize the alignment.

[nadar]

I will try your suggestion.
But doing this way, the cloud will be misoriented, and I will loose georeferencing ?

[ozbigben]

I think he suggested the export so you could import after optimisation.

[nadar]

Import after optimization to do what ?
Is it possible to use georeferenced markers to georeference WITHOUT optimization ?

[Alexey Pasumansky]

Hello nadar,

If you do not want to optimize the cloud, you do not have to. But if you are going to optimize the cloud and camera alignment for the current project, then it's recommended to remove those inaccurate markers and bring them back after optimization for georeferencing.

[nadar]

I understand that using unprecise GCP (= markers with XYZ coordinates) should be avoided for optimization procedure.
But what about markers (= with no coordinates). I can generate a lot of markers using the guided approach. On many photos, the predicted placement is failrly good, but sometimes it can be a few pixels off, and on some photos, predicted position is completely wrong.
Are these needed for optimization ? Should I use only the "trustable" ones ?

I was wondering if the origin of my problems could not be related on the fact that there are many moving objects in my images (cars, pedestrians, etc...) Does Photoscan impement a kind of a filter to discard obviously wrong stereo points ?
Should I mask these objects ? When: before aligning or during optimization ?