Platform: Website
Browser: Firefox
Screenshots of what you are seeing:
Description of problem: This occurred on about six of the 34 observations I uploaded today. The computer vision offers only these three options and none of them look anything like the organisms in the photos I took. Seriously, that plant hopper nymph is very hard to mistake for an octopus
It started giving more accurate suggestion on the same observations once I uploaded them.
Hi @isopodguy,
This bug is familiar to me. CV uses small previews for identification. In some cases, this preview is just a small black rectangle. Probably, this is due to unstable internet connection somewhere or server capacity problems.
So, CV is just identifying black picture as an underwater beast which usually captured on camera on black background.
Hope this helps.
it seems to be an issue only with firefox, I have also had it a few times
It might help to crop the photos, so: cut out the object so it takes up most of the photo. As far as I understand, the CV doesn’t actually look at the small organism in the middle of the photo while disregarding the surroundings. It looks at all the pixels and tries to find pictures in its database with similar pixels. The plant hopper nymph only takes up a small percentage of that photo; most of it is brownish cardboard (?) which actually does resemble the sea floor, I guess, so the CV will naturally come up with pictures of sea creatures.
see also https://forum.inaturalist.org/t/ai-offline-not-working/59503. Firefox has such problems if it endured a computer sleep and wake-up cycle.
Last week I had the same problem on 640 observations…it wanted everything to be a bat.
I use floorp for my browser. -
Files were coming directly from an SD card from a panasonic and olympus camera, using their standard file nomenclature.
Is the photo file named … sea s ? or bat… ?
Either way - you can start the CV with - this obs is for … at …
Seems to be the same as this:
https://forum.inaturalist.org/t/cv-giving-the-same-incorrect-suggestions-on-every-uploaded-observation/61443
That sounds like this issue:
https://forum.inaturalist.org/t/cv-suggestions-are-different-and-mainly-bats-when-uploading/25589
Apparently, this behavior is caused by a specific privacy feature called resist fingerprinting (which has nothing to do with fingerprints). I’m also wondering if it’s related to isopodguy’s problem since they are using Firefox (which Floorp is based on).
This might be the case, but it doesn’t explain why the CV always gave the same three suggestions with the same taxon photos every time.
The CV is supposed to match my photo with the best matching taxon photo, right? It didn’t seem to do this for my observations, plus a lot of the ones I uploaded had very clear close up photos of organisms (like plants) and it did the same thing for them.
Yeah, I guess in your case there is some other underlying problem. But in general, you should get the best results from computer vision models when you crop the photo and leave out as much irrelevant background as possible.
The CV does not actually see an organism in the picture like we do, automatically disregarding the background. It only sees a bunch of pixels and tries to find the taxon photos with similar pixels. So, if I’d upload a photo of a flying dragonfly against a blue sky, where the dragonfly takes up only a small portion of the photo, you and I see a dragonfly, but the CV does not actually see a dragonfly like we do; it just sees a dark dot on a blue background. And it would then probably suggest that it’s some kind of hawk or other bird, because the database contains thousands of similar looking photos of a small dark dot on a blue background, where the dark dot is identified as a hawk or other bird. Now if I would crop the photo, so the dragonfly takes up most of the frame, the CV focuses less on the blue pixels and more on the shape of the dark pixels, which form a long body and four wings, which does resemble photos in the database of dragonflies. So it’s more likely to come up with a correct ID.
Thank you so much guys!
@ [thomaseverest] @ [zygy]
I’ve seen CV identify others’ unknown observations as mallards or other birds, when they clearly aren’t. I’m using Brave. I don’t know what browser the observer used; observations that are uploaded without ID are more often from one of the apps.
