SE++ processing fades out

[mkuemmel]

Following the discussions this morning I have uploaded the 21GB dataset to here: https://deepdip.iap.fr/#item/62332d6381e19044840fe096

The two shell scripts in there both do not run, which means they fade away after the background modelling. Note that:

• I have done a fresh compilation of SE++ v0.16 and v0.17;
• both v0.16 and v0.17 show the same error, meaning the problem existed already in v0.16;
• looks like there is a dependency on the dataset, since v0.16 and v0.17 run on the very similar Euclid dataset from last year;
• the dataset does run with v0.17 on a SDC test machine, which is a native EDEN.3.0 machine;
• I am running on a Ubuntu20.0.4, levmar is installed from the sources, SE++ is installed via Elements and ElementsEnv

Could be related to #445 though some modifications have been done but it does not help in this case.

[ayllon]

Hi Martin,

At some point both model coordinates (x and y) are NaN, which on a debug builds triggers an assert error but on a release build likely keeps going until something else triggers another error that kills the thread.

It may be somehow related to the bug on the iterative model fitting that @marcschefer mentioned yesterday, since when I go back to the "bulk" model fitting with

use_iterative_fitting(False)

It runs just fine.

[mkuemmel]

Thanks for checking that so quick! I can confirm that switching off iterative fitting solves it.

I was wondering by myself whether the iterative model fit bug causes this problem, but in principle all detected objects are on the measurement images.

I just checked, with #462 iterative model fitting can be switched on again and seems to solve the problem.

[ayllon]

I think we can close this one?

[mkuemmel]

Right, we can!