Feature Reprojection Error Threshold?
AnsweredI couldnt find a clear answer to this so trying to ask specifically:
When I set the Feature reprojection error threshold to 0.2 for instance, i still get a median and mean error above 0.3 to 0.4 in the alignment report. Isnt the threshold supposed to rule out alignments above that error? I know it would eventually result in more components and less images being taken into account, but how could I throw away the less accurate images?
Thanks
Daniel

Dear Daniel,
we need to check it out. Maybe there is some minimum inside the algorithm that is not reflected to the application settings. thank you for reporting.
Anyways, I would say that it makes no sense to set the reprojection error to 0.2 ... I would say that minimal reasonable value is 1.0 ... maybe more. But I need to check it definitely. 
Thanks so much Michal. So far my experience is that whatever I change in the threshold, 0.2, 0.15 or 0.1, the report is the same, median and mean errors around 0.34 to 0.4..
My aim is to really push the quality and eventually get rid of images that dont align too good.
The alignment time increases extremely though when going lower like 0.2 or even 0.05.. so something does change, at least in the calculation..
Looking forward to the outcome of this!

I'm suspecting that mean and median error are not a measure of Alignment quality, but simply the mean and median of the longtail distribution curve that results from any large sample of points, ties or whatever  and that curve is pretty much the same every time. Those typical 0.3 to 0.5 numbers seem to remain in the same range under all Alignment settings, with two exceptions 
One is with very poor small photo sets, where there's a significantly small no of points, ties etc, so the distribution curve differs from the well fixed largesample curve;
The other is under influence of Max Reprojection Error, which applies a topcut to the long tail of the curve, hence mean and median shift close to the peak of the curve.
Think of it as a set of four figures  minimum, mean, median and maximum (minimum in all cases being zero or v close to).
So, really quite uninformative. The need to get Alignment mean and median error under 0.5 is meaningless, as it's always that under almost all circumstances.
True or false?

Now you've challenged me to sharpen up my vague intuition! I realise it's not a 'longtail' curve as currently spoken these days (e.g. in https://www.amazon.co.uk/MoveFastBreakThingsFacebook/dp/1509847693/ref=sr_1_1?ie=UTF8&qid=1519380181&sr=81&keywords=move+fast+and+break+things )
I had in mind a curve shape like https://getrevising.co.uk/https_proxy/6223  ignore the writing, just the curve shape.
The x (horiz) axis is the error and the y (vert) axis is the count of tie points (TPs) at given error. The full curve would be seen if Max error is set very high. Decreasing Max error is a vertical line that cuts off the right hand end of the curve.
Mean error is a vertical line that's going to lie v close to the curve's ypeak and the Median error is a vertical line somewhat to the right  higher x.
Decreasing Max error (moving its vertical cut off line to the left) pushes the other two vertical lines to the left. At very low Max error, Mean error may fall leftward off the peak and decrease rapidly down the slope. Mean error will be more resistant to decrease  may in fact increase a bit as it's pushed leftward to the peak.
Increasing Max error (moving its vertical cut off line to the right) pulls the other two vertical lines to the right but causes a diminishing increase in Mean and Median error.
Does this correspond with reality? Seems to, in the numerous cases I've run. Whether they're good or poor, larger or smaller photo sets, the Mean and Median error figures seem nearly same in case after case, regardless of Alignment settings, only affected, as I say, by very poor small photo sets, and by the Max error setting.

Hi Tom,
wow.
You do have a background in mathematics, right? I've tried to get the definition of mean and median error but there were only equations and i got bored. Could you explain the two to normal people? ;)
Anyway, it does make sense the way you explain it. That would mean two things:
1. The quality of the alignment is determined by the images and not by the Max Repro Error.
2. The Max Repro Error is only there to basically filter the worst points out of the cloud, which would impact the meshing more than anything else.
This all goes with a loss of the overall number of tie points, for good or bad. It would be great to have that curve available, so that one can estimate at one glance where to lop off the top.
Apart from that, there is one thing I want to add. The errors are definitely influenced by exif grouping. With grouping on, they can be worse (especially with change of focus and/or aperture at the same focal length) but the chance for a proper alignment and all or most images in one component is much higher. Switched off again, it usually decreases by about 20% in my case, which is due to slight adjustments that can then take effect.

Alevel (age 1617) school Maths Phys Chem only  and Maths/Algebra I really struggled with  except when I could visualise relationships as a graph.
Mean is what is normally understood as 'average' i.e. with 5 observations of 2,5,7,11,15, mean is (2+5+7+11+15)/5 = 8.
Median is the middle observation i.e 7 (but I seem to think, in a tighterspaced set of observations, would be a bit larger than Mean).
Götz Echtenacher said "so that one can estimate at one glance where to lop off the top". To be clear, when you say 'top', you mean 'right hand side'? You make it sound great  what exactly would 'one glance' tell you?
Last para v interesting, and does that 20% mean "better, both in 'most in one component' and in errors" by means of having EXIF grouping first on then off? Encouraging RC in its (sometimes) "improving Alignment over a series of runs" tendency?
Please sign in to leave a comment.
Comments
10 comments