Description
Description
While doing user testing on Centaurea melitensis, I found that the seed detection model had a REALLY hard time detecting this species properly.
Bug Summary:
For some reason, the model had a really hard time detecting the correct number of these seeds. I had to adjust the positioning and then increase the contrast to the point where the pappus (the white fluffy part at the top) pretty much disappears.
Steps to Reproduce:
Load any of the images below and classify. The results in my screenshot were from the 27 species model but I don't think the seed detector has changed?
Expected Behavior:
Five seeds would be detected, with boxes that overlapped if necessary.
Actual Behavior:
Boxes either straddle multiple seeds or some seeds are missed entirely.
Environment:
Commit Hash: a427278e-28df-428f-8937-ddeeef44e72f
OS Version: Windows 11
Browser Version: Edge
Extra Screenshots:
This last picture was finally detected properly, but you can see it is pretty strange-looking compared to the other ones.
Additional Context:
I suspect the problem is something to do with the pappus (the white feathery stuff at the top of the seeds) having low contrast to the background and not a strongly defined edge. When taking images of seeds with pappus we often use a darker background, but I worried this might mess with the results further since the model, as I understand it, was trained with light background images. On the up side, the IDs were all correct!
Suggested Fix: (facultatif)
[If you have an idea of how to fix the bug, describe it here.]
Metadata
Metadata
Assignees
Labels
Type
Projects
Status