Computer vision suggestions way way out of native range

General
1

Not sure of the best category.
I just reviewed a post from Kyrgyzstan showing a photo of a juniper and the ID was Juniperus grandis, which is native to California and Nevada in the US.
https://www.inaturalist.org/observations/251288646
I’m guessing that computer vision offered this suggestion.
I posted the ID of Sabina (scale-leaf junipers) then looked at suggestions from computer vision to give me clues where to look in Junipers of the World. It offered 4 suggestions, 2 which were likely, one nearby but unlikely, and Juniperus virginiana, which was as off base as J. grandis.
Are there no regional limitations to CV suggestions?

2

CV now offers (me) your choice of pseudosabina

3

The CV does take location into account when available. See this blog post for more info about that. The problem is that it’s also perfectly happy to make suggestions when location isn’t known (maybe the user hasn’t put it in yet, or metadata from the photo is still loading when they get the CV suggestion). Why this is still allowed eludes me, since it’s one of the top source of misidentifications on the site.

4

It is a CV ID posted with Seek. I believe Seek uses an older and simplified version of the computer vision, which may not consider the geomodel/expected nearby in the most updated version either. The current website model is more recent. With the website model, the CV is ‘pretty sure’ it is Cupressaceae, with top suggestions of Juniperus pseudosabina and Juniperus sabina, with a strong preference for Juniperus psuedosabina between those two options. Turning expected nearby off, it also has Athrotaxis cupressoides and Phoradendron juniperinum ahead of Juniperus grandis.

Unfortunately seek IDs are just going to be less reliable at present.

3 Likes
5

I’m pretty new here, but it has been my perception that:

  • when I submit observations
  • using images without embedded location information
    (workflow = digital camera without GPS –> iPhone Photo app –> crop & comment –> iPhone iNat app –> upload)
  • if I ask for CV suggestions
  • before manually providing a location,

…then CV suggestions have no regard for my location.
If I accept a CV suggestion and then enter a location, there is no guardrail to pull back the suggested ID which is now clearly out of range.

If that is actually the case, it seems this would be more a workflow/app issue than specifically a Computer Vision issue?

6

Seek does not take location into account when providing ID suggestions. And it’s using a very old version of the computer vision model.

We’re currently testing a version of Seek that uses the same model version as iNaturalist Next (2.13, I believe) for live ID suggestions, but it will take some tweaking to get the app behaving properly. Eventually we would like to also incorporate the geomodel in the live ID suggestions, but with larger models and then the geomodel, it makes for a bigger app and potentially slower responsiveness. Finding a balance that provides accuracy and is still responsive in use is difficult.

2 Likes
7

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.