Apparent bug in blob pruning code #2022
Conversation
…eviously already been pruned.
skimage/feature/blob.py
Outdated
| @@ -92,6 +92,13 @@ def _prune_blobs(blobs_array, overlap): | |||
| # iterating again might eliminate more blobs, but one iteration suffices | |||
| # for most cases | |||
| for blob1, blob2 in itt.combinations(blobs_array, 2): | |||
|
|
|||
| # Ignore if we have removed either blob during an earlier iteration. | |||
| if blob1[3] == -1: | |||
There was a problem hiding this comment.
if (blob1[3] == -1) or (blob2[3] == -1)):
There was a problem hiding this comment.
But I've just realised that it should actually be
if (blob1[2] == -1) or (blob2[2] == -1)):
Because the 'radius' is at index 2. Apologies.
|
The code seems to do the radius detection, but I don't see the first step that you mention in the PR description? Also, I want to double check: this is an optimization, not a bug, right? |
|
Hi Stefan. The situation I had can be described in this example. In the original code, say two blobs, A and B are compared for overlap and let's assume B is correctly marked for removal. This is done by setting its 'radius' to -1. Later B is compared with C but their overlap is now measured assuming that B has a radius of -1. This can give an incorrect result and lead to C being incorrectly marked for removal. |
|
@evariste the change looks good. Would you mind creating a test to cover this case? |
When pruning blobs that have overlap, some blobs can be removed when comparing with a blob that has previously been removed. Also, blobs with a negative radius, set in an earlier iteration are not passed to the overlap calculation function.