New Computer Vision model released!

Fair enough, thanks!

As an aside: from my understanding, “research grade” observations are required for the training, which are very rare for microbial observations. This is for two reasons. 1) very few people are able to identify microbes with any degree of rigor and 2) as I already mentioned, few legitimate IDs are at the species level. I think I read that it’s possible to mark an observation in such a way as to remove the species level requirement from research grade, but this is such an unusual operation that we cannot expect it from most users. So, I suspect the machine suggestions for microbial life will unfortunately continue to be essentially useless. Such is life! Fortunately, it is not a big concern for me, only a “nice to have someday.”

Yes, you can check a box in the DQA that the community ID cannot be improved, and the observation will become research grade at the level of the community ID.


provided the ID is at below family


Yes, I think they were referring to genera anyway, though.

Hopefully once we figure out how to display it in a way that increases clarity without adding confusion, we’ll add a label to indicate when genus or family is in the model because none of its children are. We just haven’t quite worked that out yet.

In the meantime, if you have a question about a specific taxon or group of taxa, I can look it up for you.


Is there any way the CV could be trained on just feathers?


Probably, but it would probably need to learn how to sort the feathers in the categories, like tail, wing, contour, down, etc., to make it easier to identify.

1 Like

I think the CV doesn’t take life stages of animals into account, so maybe it doesn’t necessarily need to know that? It would still be interesting to try it even if it couldn’t tell what type of feather it is.

1 Like

People submit photos of feathers as evidence and they end up in the training data, but I think a dedicated model just for feathers would probably do better. If you’re interested in trying to create something like that, keep me in the loop! I’d love to know how it goes.


I wish I could, but I have no idea how to. I’m also not very good at programming.

Is there a plan to have subspecies in the CV? For example, there are currently 26965 of sambucus racemosa (red-berried elder) but 101 observations of sambucus racemosa melanocarpa (rocky mountain elder), which is a regional variant which is black. If you are randomly selecting 1000 photos for the CV, at most 3 or 4 of those will be melanocarpa, and some of them will show things like the leaves and not the actual black berries. Consequently, in my region the CV usually suggests something not found locally for an elderberry which is black, even though melanocarpa is the only black elderberry in this area. Even if it suggest racemosa based on the leaves, a user inexperienced in the taxa might scroll through the first few photos and decide it can’t be that because seemingly all of the photos have red berries, and click something else. As melanocarpa is now over the threshold where it could be included in CV if inaturalist taxonomy assigned it as its own species, it seems like it would be useful to train it as one.

I’m sure there are plenty of similar examples in other taxa.

See for opinions for and against this idea.

3 posts were merged into an existing topic: Fail Whale on Certain Links

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

now that hybrids are excluded from computer vision suggestions (see, it’s causing a recent unfortunate string of misidentifications. Crocosmia × crocosmiiflora is a widespread/common weed in Australia, with a lot of observations. When hybrids were in the CV, this species had been in it for ages due to the high number of observations. Unfortunately now that it’s out, Crocosmia aurea, which is not recorded from Australia, keeps getting offered as a CV suggestion instead. There have been ~15 of them pop up for NSW thus far, so not a huge amount, but a little annoying


(moderators, feel free to move this to a different topic if you wish. Very few are open right now concerning CV misidentifications.)

I just had the CV very confidently suggest Urocerus albicornis for a somewhat blurry photo of an ordinary fly. I thought I should share my anecdote, since there have been some discussions in the past regarding such misidentifications.

1 Like

While this thread is open… I’ll mention that the exclusion of common hybrid plants such as Crocosmia × crocosmiiflora from CV training continues to cause problems, as lots and lots of wild observations worldwide are automatically misidentified as Crocosmia aurea. I get that there is a problem with avian hybrids, but excluding plant hybrids appears to have caused another problem for accurately tracking invasive weeds.

Without the prospect that CV might be able to supply correct identifications in future there’s not much incentive for identifiers to play whack-a-mole by correcting IDs for invasive horticultural hybrids. @alex: Would it be possible to add plant hybrids back into CV eligibility?


And also while the thread is open …
I think it is worth discussing what happens around species that are part of a species complex. By definition the members of the complex should be morphologically similar, so it would make sense for the associated suggestion to be bumped up to the rank of species complex and not one of the subordinate species. I guess one response is that members of a complex often have restricted ranges, but the new geomodel is quite poor at identifying those ranges, at least for my group (the fungi). As a consequence the CV often suggests species within a complex that aren’t loca/within range. Bumping up to complex would be correct more often, and could then be narrowed by the application of real intelligence - by humans.


Hi Rupert,

We’re got this identified as an important upcoming experiment and I’ll let you know when we have some results.