recommended value for reprojection error limit

[jeremyeastwood]

I would like to remove all points with high reprojection error after the align stage to improve my model quality - any recommendations for a sensible limit I should use?

I had been using a limit of 1.0 in photoscan 1.0.4, however in 1.1.0 this value now seems to be too aggressive (cutting out most of my data), so I was wondering if there is a specific value which indicates very inaccurate data that I could use for all my models?

Thanks

[bigben]

I typically use 0.5 for reprojection error and 35 for reconstruction uncertainty. I save the project before applying these so that I can always go back and choose higher values if it removes too much.  It can be a bit tricky determining what "too much" is, but a) optimising the cameras afterwards will throw up a warning that some cameras have insufficient tie points and b) at the start of building the dense point cloud the console will show the number of images selected to build the point cloud.  I've had projects where 80% of the sparse cloud was removed but it still selected 90% of the images for the dense cloud....

Getting a good calibration for your lens goes a long way towards getting consistent quality, especially as you go to a wider fov. If you have the lens calibration set to auto you can improve the result by saving the adjusted lens parameters, then changing the initial setting to precalibrated and load the lend detting you just saved. Rerunning the alignment after that will refine the lens parameters further and usually improves the result.

[jeremyeastwood]

Thanks for the reply - very helpful!  Strange why removing points with reprojection_error > 1  is cutting out so much of my model then (cutting the points out ends up producing large gaps in my dense cloud... ).

Thanks for the tip on lens calibration - do you normally find that one extra align seeded with the adjusted calibration parameters does the trick, or do you keep iterating (and if so, which parameter are you trying to minimise?)?  Weird side issue I'm having with this step: I can save the adjusted calibration params as an xml from the gui just fine, but when I try to load them back in to the initial values, I get an error message with "could not load parameters" - is this the same procedure you use, or am I doing something wrong?

Thanks again for your help - it's really appreciated

[bigben]

If the distortion parameters are off then there will be larger errors.  I'm guessing the calculation of the distortion parameters is similar to that for panoramic stitching in that it's essentially a mathematical "guessing" algorithm. How quickly it gets to the right values depends on the starting values and the data (images) it has to work with, but there'll be a limited number of repetitions programmed in so that you don't have to wait forever to get a result. Repeating the process is essentially just extending the time you wait while the values are further refined.  But depending on the scene and your camera positions, it's also possible that you may not get a great result.. In those cases it really helps to have a good lens calibration.

Agisoft lens is a good starting point, although I usually use a specific scene for a slightly more agricultural approach... mainly because I'm looking at wider lenses which makes filling the image with a monitor quite difficult.