Non-sequential data in supervoxels during training can result in errors during learning, as well as log(0) errors (due to non-existing segments). See this thread on the mailing list for details.
The simplest fix is to run skimage.relabel_sequential on both the superpixel and ground truth volumes, but I'm not sure whether other things would break at that point (and we might need to maintain the inverse maps if specific labels need to be reported back to the user). Ideally, we would have a test case to ensure this doesn't regress.
Non-sequential data in supervoxels during training can result in errors during learning, as well as log(0) errors (due to non-existing segments). See this thread on the mailing list for details.
The simplest fix is to run
skimage.relabel_sequential
on both the superpixel and ground truth volumes, but I'm not sure whether other things would break at that point (and we might need to maintain the inverse maps if specific labels need to be reported back to the user). Ideally, we would have a test case to ensure this doesn't regress.