Transfering parameters from "tracking with learning" to "automatic tracking" in ilastik

ilastik
#1

The tracking with learning workflow has calculated tracking weights for me but does not allow batch processing. I would therefore like to transfer these weights to the automatic tracking workflow, but the scales are quite different (e.g default for division: 0,6 in twl, 10,00 in at). Is there a factor to multiply by so that I get the same results via both workflows?

1 Like

#2

I believe you’ll get the same results given the same weights. If that is not the case, please let us know, we’ll take a closer look. But just re-using weights is the intended strategy.

0 Likes

#3

Hello,
sorry it took so long to respond. Twl gives me a weight of 0,72225 for appearance, but at does not accept values between 0 and 1, so I cannot transfer the weights directly. Also a weight of 1 seems quite low for appearance

0 Likes

#4

It took some time to dig out, but now we know. The detection weight parameter in the automatic tracking workflow is always set to 10. So you need to rescale the values you receive from tracking with learning so that the detection weight is 10. That should give you the same results in both.

Hope this works, let us know and we will update the documentation meanwhile. Thanks for the question!

1 Like

#5

Thank you, that is great. However, right now it gives me a detection weight of 0,0000 (before that it gives a negative detection weight). How should I proceed?

0 Likes

#6

Hmmm, I’d say make it small and positive, like 1, and everything else much bigger than 1 and in the right ratios to each other. But it is a bit pathological with detection weight 0. This usually happens when the object classifiers are not providing sufficiently clear indicators to the main graphical model. I wonder if it can’t be pushed off this parameter point by a little bit more labeling? Could you try to go back to division and object number classifiers (save your current version in a copy) and enable the uncertainty layer (you can do it in the lower left layer stack widget). Look out for objects of very high uncertainty and label them with the correct class. Once the uncertainty level in both classifiers has gone down, try to compute parameters again.

0 Likes

#7

ok, thank you, I’ll try that and also label more objects

0 Likes