How to keep only the best data?

[maddin]

What is the recommended strategy to keep only the very best data in an aligned project?

Let's say I have a project with many cameras but the overall result (dense cloud, mesh) isn't great. So I would like to have some systematic way of pruning photos/cameras/points/markers from the project that are less reliable than others. The result would probably something with less coverage but hopefully those parts covered are really nicely reconstructed.

To start with, we have the 'Image Quality' that somehow measures the sharpness of the sharpest part of the image. Is that a good measure and what threshold should I use?

I also know about the Gradual Selection tool but those unitless thresholds seem very arbitrary, too, and don't tell the complete picture each on it's own (e.g. I could throw away all points above a certain reprojection error, but then maybe the angle between the cameras isn't large enough so their 3D location is rather weakly defined (this is covered by Reconstruction uncertainty IIRC)), so I am looking for help how to combine these. Also, Gradual Selection does not apply to the dense point cloud, does it?

Then there are some reprojection numbers in the Chunk Info, as well as some error numbers for each camera in the Reference Tab ('Error (pix)'). 

So now the question is how to combine these all into one consistent approach to only keep data that looks very reliable?

Thank you for any insights

[Thibaud Capra]

A good threshold for the Estimate Camera Quality results is to remove all cameras with a score inferior to 0,5 according to the PhotoScan User Manual.
From my own experience, I've had even better results when removing photos under 0,6 quality score. by doing this you have to watch out for sufficient overlap though!

Using the GSD you should also be able to have an a priori estimation of your accuracy, then compare it to your end result to see if you need to refine things.

(I'll keep an eye on this topic it might be very interesting!)

[maddin]

A good threshold for the Estimate Camera Quality results is to remove all cameras with a score inferior to 0,5 according to the PhotoScan User Manual.

What does that number stand for? What is the difference between 0.5 and 0.6 in terms of picture content?

[Thibaud Capra]

As far as I know, from what I found on Google Scholar etc., it's an arbitrary scaled-number that's supposed to reflect the sharpness of the images.
Let Q be between 0 and 1, 0 being the blurriest image possible and 1 being a sharp, steadily-acquired image.

See Shervais, K. & Dietrich, J., 2016. Structure from Motion (SfM) Photogrammetry Data Exploration and Processing Manual.

Also depending on your project, using the aggressive depth filtering when generating your point cloud might lead to cleaner results.