Located beads are not valid if the Airy fit is below threshold even if the Guassian fit isnt.

[iandobbie]

With the dev branch now having both airy and gaussian fits which fit decides how many valid beads there are? We have a SoRa superres (maybe, lets not argue that here!) dataset that has relatively low s:n and bad airy fits. When passed through the R^2 thresholds of 0.85 for both fits we get back no valid beads and so no data is saved. If I set the airy R^2 threshold to 0.5 and the Gaussian to .85 then I get a bunch of valid fits and output data is saved.

The long term solution is to go with whatever we are using for the final fit, or toggle between the two. Short term I think the as the Gaussian fits have much higher R^2 we should use the Gaussian to make valid/invalid decisions.

[iandobbie]

Just as an example, this is a farred bead set from out Nikon Sora fitted as suggest above.

<style> </style>

Fit Airy R2 Z Mean	Fit Airy R2 Y Mean	Fit Airy R2 X Mean	Fit Gaussian R2 Z Mean	Fit Gaussian R2 Y Mean	Fit Gaussian R2 X Mean
0.962	0.678	0.629	0.987	0.996	0.996

So the Airy xy fits are about 0.65, but the Gaussians are 0.99

[juliomateoslangerak]

I realized that, indeed, I have to bring the airy thresholds way down. I guess the easyest is to set the default value of the airy fit way lower. to 0.4 for exemple... Is the default 0.85 gaussian fit threshold ok?

[iandobbie]

They seem to be. I regularly get .95 and above, often near 0.99, so yes I think .85 is pretty generous. We may increase it, but we haven't yet consistently gone through our data. I plan to start this next week, to try get finalised data for our paper.

[iandobbie]

I suggest that we just change the default airy fit threshold to something much smaller like 0.4 or similar. A better solution would be to improve the LoG fits, but this is a longer term goal.

[iandobbie]

I just set my airy threshold to 0.4 and ignore the fits. I will close this as I think it is really part of the bigger issue of what we should use for fits. I agree that the airy fits are more correct, but I think that the fitting is much less robust and very often generates terrible fits.