The AI mistakenly suggests an American species (Rainieria antennaepes) for a European one (Rainieria calceata) half the time. They’re indistinguishable at the kind of resolution typical of iNaturalist so the AI will never learn unless a geographical component is introduced.
When can we expect this?
Does anyone have other examples?
1 Like
a geographical component already does exist. When selecting AI suggestions from the dropdown, there’s an option to select ‘only view nearby suggestions’. Of course, this can still be influenced by pre-existing misidentified observations nearby, but it cuts down on these cases quite a lot
(also I’ve moved this to general since it’s not a bug)
5 Likes
That’s not much help to verifiers as that feature is ignored by half the users posting images.
Do verifiers have an equivalent button to press. Or a macro which will save us having to make the same comment in the “Tell us why” box
It’s a recent addition, so old posts weren’t affected by that and cv suggested species all over the world, it’s not ignored now when it’s automatically applied and new users won’t ignore it if they rely only on cv.
4 Likes
I use Text Blaze.
3 Likes
How old are the observations you’re coming across? As has been noted by others, as of a few months ago the computer vision suggestions only show “seen nearby” results by default so this shouldn’t be happening much anymore, so it would be interesting to see how much of a problem this still is. What we need to do as a community is clean up the falsely identified observations so they don’t show up as “seen nearby”.
5 Likes
Today - https://www.inaturalist.org/observations/88245299
A day or so ago https://www.inaturalist.org/observations/87439079
https://www.inaturalist.org/observations/86990208
July 5th https://www.inaturalist.org/observations/85824188
If you need example stats then my project at https://www.inaturalist.org/projects/european-micropezids-tanypezids will have them, just filter for all the R. calceata records. That is the only one which does this within my project.
I’m not sure if this is quite the same issue, but since you’re examining the issue, I’ll offer an example. I was looking at this observation of a moth in Canada. The AI/computer vision suggestion is supposedly on “include suggestions not seen nearby”, as shown in the attached image, however at the top it gives “We’re pretty sure it’s in this genus”, however that genus is only found in Australia as far as I can tell:
Perhaps I’m misinterpreting the suggestions, but it’s confusing to see a genus from so far away when suggestions are apparently only being shown from nearby.
3 Likes
https://www.inaturalist.org/observations/88245299 and https://www.inaturalist.org/observations/87439079 were made via the Android app and https://www.inaturalist.org/observations/86990208 was made via Seek.
Seek does not take location into account when suggesting an ID, which explains that ID, and we believe the Android IDs stem from a recently-discovered issue in Android, which we’re working on. Testing the photo (with the same location and date info) from https://www.inaturalist.org/observations/88245299 in iOS and the website doesn’t return Rainieria antennaepes.
For https://www.inaturalist.org/observations/85824188 - it was uploaded via the web but I think the issue there is that the photo isn’t cropped well. Take a look at the text here:
Basically the model didn’t find anything particularly visually similar and couldn’t find a common ancestor so it it just showed top suggesions without any filter. The idea is that it’s better for there to be an ID of some kind so it will get noticed and corrected if need be.
But note that it was uploaded when our last vision model was still in use and we can’t really test that directly anymore.
This is a known thing. That ID is relying purely in visual similarity it’s not using any location information.
5 Likes
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.