Closed JaimeIvanCervantes closed 8 years ago
This definitely reduces the memory footprint, but it means relabelMergers()
is run twice per frame.
Unfurtunately, fitting GMMs (our current merger relabeling technique) is not deterministic, and so one might get out a tracking solution that doesn't necessarily fit to the images.
We could add another method to our merger resolving plugin and use the fit that we have already computed and stored (we only use the fits of the previous frames here https://github.com/chaubold/hytra/pull/6/files#diff-060028ca83bd4e2a8b4f1914aed823f5R137).
By doing so, we could simply not call GMM.fit()
in https://github.com/chaubold/hytra/blob/development/hytra/plugins/merger_resolver/gmm_merger_resolver.py#L39 but only call GMM.predict()
to reassign the pixels.
The
relabelMergers()
function was added, and is called from withinopConservationTracking
in Ilastik to request frames with refined merger segmentations.