Add .correct_mask method
#27
Conversation
HaleySchuhl
added
work in progress
|
Here's the code health analysis summary for commits Analysis Summary
Code Coverage Report
|
check against overlap with any annotation rather than just considering pts auto detected somwhere vs kept coordinates after fixing annotations
minor context update to example use section
watershed can be used upstream to separate conjoined things
…ed mask is handled
There was a problem hiding this comment.
Everything seems to work as intended. I put the notebook I used for testing in the Google Drive folder. It seems like there's just a test coverage issue and one deep source issue to fix. Once the coverage is good to go, I can re-run the tests and approve though!
|
I'm not certain we will be able to decrease the complexity, which is the deepsource complaint. But coverage is fixed now @annacasto at least! |
plantcv/annotate/classes.py
Outdated
| for names in labelnames: | ||
| for (x, y) in self.coords[names]: | ||
| x = int(x) | ||
| y = int(y) | ||
| # Draw pt annotations onto a blank mask | ||
| pts_mask = cv2.circle(pts_mask, (x, y), radius=0, color=(255), thickness=-1) |
There was a problem hiding this comment.
Take a look at itertools, particularly itertools.product I think to reduce nested for loops to one loop. This also applies to dealing with complexity below.
There was a problem hiding this comment.
Thanks for the feedback! I have simplified this helper function.
I spent most of this morning testing different ways to loop through all annotation coordinates, but I'm struggling to find a way to avoid the nested for loop within the actual correct_mask method. The part that I'm not sure how to get around is that the algorithm currently relies on being aware of the corresponding class label called names, but I believe it's possible. If you have more suggestions I'm happy to implement them, and I'm also amenable if you want to make changes directly.
There was a problem hiding this comment.
I have removed the nested for loop with a different approach (edited) butitertools.product is the first way I tried this, and I'm just not sure how to use it in such a way that keeps awareness of the key part of the dictionary rather than just gathering all coordinates. Here's the closest I got today:
import itertools
# still a bad product
for instance in itertools.product(counter.coords.keys(), counter.coords.values()):
print(instance)
('default', [(398, 416), (1262, 338), (1701, 378), (1806, 1025), (1518, 1505), (556, 1430), (481, 997), (1090, 754)])
('default', [(321, 486), (308, 888), (333, 993), (321, 1193), (338, 1489)])
('default', [(1423, 529), (1721, 399), (1864, 229)])
('germinated', [(398, 416), (1262, 338), (1701, 378), (1806, 1025), (1518, 1505), (556, 1430), (481, 997), (1090, 754)])
('germinated', [(321, 486), (308, 888), (333, 993), (321, 1193), (338, 1489)])
('germinated', [(1423, 529), (1721, 399), (1864, 229)])
('other', [(398, 416), (1262, 338), (1701, 378), (1806, 1025), (1518, 1505), (556, 1430), (481, 997), (1090, 754)])
('other', [(321, 486), (308, 888), (333, 993), (321, 1193), (338, 1489)])
('other', [(1423, 529), (1721, 399), (1864, 229)])
I found that this sum syntax I've added to _create_pts_mask does a nice job of combining all annotations into a single list, but for the algorithm to utilize this format of data, we would also need to create a list of the same length that contains the class label corresponding to each value. (I'm currently doing this with a for loop over all annotations and looking up the key for each value)
I'm just not certain that itertools.product is the method that can create this based on the googling & testing I was doing today, but that could be my lack of familiarity with the package. So if somebody else wants to replace the logic then feel welcomed to directly commit those improvements.
Describe your changes
Rewrite the
clickcount_correctmethod, used to add and remove to make acorrected_mask, which makes a labeled mask containing all objects that are resolvable in the input binary mask, based on Point annotations. Outputs can be used for size analysis. Unresolved objects cannot be measured but can still keep info about count.No longer includes the
_recover_circpart of the algorithm which was changing the shape of objects in some cases and was too specific to the size & shape of pollen.Type of update
Is this a:
Associated issues
Reference associated issue numbers. Does this pull request close any issues?
Additional context
Add any other context about the problem here.
For the reviewer
See this page for instructions on how to review the pull request.
plantcv/mkdocs.ymlupdating.md