Multicut agglom skips labelling neurons even when the watershed labels exist

[Mohinta2892]

Hi!

I am running the example multicut.py script on a 256x256x256 block to test distributed processing.

I also have a precomputed watershed, hence I pass (my affinities also need to be inverted):

    run_mc(path, tmp_folder, max_jobs, target=target, have_watershed=True, from_affinities=True, invert_inputs=True, n_scales=1)

Multicut agglomerates the watershed, however, it skips assigning a label to one neuron as you can probably see from the screenshot.

Corresponding watershed:

I am not sure what is going on, so any help is greatly appreciated!

Thank you!

[constantinpape]

Hi @Mohinta2892 ,
the reason for this is that one of the segments is assigned the ID 0, which is then not displayed in the viewer. It depends on your further analysis if this is actually an issue or not. (If 0 is treated as some extra ID corresponding to background then it is an issue and should be fixed).

If your volume is small enough (as for your example data) you can just load it into memory and add 1 to the whole segmentation.
If that becomes difficult / inconvenient due to the volume size: we could also fix this in a more principled way, by adding the 1 here. These are all the segmentation IDs before mapping them back to the volume.
I think it would make sense to fix this on that level since having 0 as a reserved ID for background is quite common.

[Mohinta2892]

Hi @constantinpape,

Thank you for explaining, I understand now.

It is worth fixing the script because the idea is to run this on a massive dataset. I am trying to scope out all failure modes by running on a test set first.

I will add the 1 and rerun the tests.