High percentage of detection misses

Bug Reports
1

In which Wildbook did the issue occur? ACW

What operating system were you using? Win 10

What web browser were you using? latest chrome

What is your role on the site? admin

What happened?
I’ve been working on deleting 18 sightings from ACW, that need to be corrected and re-uploaded. As I was deleting each encounter, I noticed something I’ve been noticing in recent days as I’ve been looking at the encounters for this species. There seems to be a very low/poor detection rate, particularly when there are multiple individuals in an image, but not only. I saw one earlier today that had an animal, standing, full right profile, in the foreground of the media asset, but the bounding box was of something indistinguishable way in the brush in the background.

What did you expect to happen? much better detection accuracy and completeness.

What are some steps we could take to reproduce the issue? I’ve deleted the encounters that I have examples of but I’ll email a doc to services@wildme.org with screenshots of the more egregious examples I saw in this small sampling of sightings.

2

Hi Maureen,

Thanks for letting us know. Because the detector is running as evidenced by the other bounding boxes, these may be legitimate misses by machine learning.

Until we have time in the future to annotate more bounding boxes and retrain the detector, I recommend using the Manual Annotation tool as a workaround.

Thanks,
Jason

3

Hi @jason, I understand what you’re saying but I’m not sure how feasible that solution is given the current limitations of the Add Annotation feature, and the thousands of images affected, in this one census alone (we have another census pending upload and this year’s census is about to begin).
It sounds like we probably need to expedite re-training on this species. @PaulK - FYI

1 Like