I was uploading photos today on the website and I noted that the drop-down CV suggestions for multiple photos would give no suggestions at all, just the text “We’re not confident enough to make a recommendation.”
It was only a handful of photos (I included most of them in the screenshot above), but I don’t think I’ve ever had the CV not recommend anything in all the time I’ve been using this site. I feel like I’ve uploaded comparable photos to those above (and worse!) and always gotten a drop-down of possible suggestions.
Bizarrely, I sent the offending photos back to my iPhone to try uploading them on iNat Next. What I got was this screen
Now, that second screen is what I’d expect of the CV for a distant picture of a pine tree, but apparently that’s only using offline suggestions for some reason?
Having recently been identifying unknowns in my area, I’ve come to suspect that this is not an issue with the web photo upload or with my own account, but potentially with recent changes to the CV (perhaps the recent photo similarity update?).
Here are three records for which the CV gives no suggestion at all, and for which I would have expected it to in the recent past:
(That last record does now give ID suggestions, but only because I have suggested an ID; if you go to the ‘Compare’ pop-up on the ID and clear the ‘Taxon’ filter, it once again shows ‘No Suggestions’)
In fact, if one opens one of the first two records in ‘Identify’ and uses the ‘Suggestions’ tab, it likewise shows ‘No Suggestions’:
Has anyone else encountered this? I don’t remember the CV ever being completely unable to give a suggested ID—and especially not for good photos of plants with flowers and leaves—but maybe I am just severely misremembering!
i doubt that this is new. since a change last year to evaluate non-square images using a center square crop, there’s always been a greater potential for unexpected results, if your diagnostic features fall outside or on the edge of the center square.
i think there was also some leftover increase in the inconsistency between how the web uploader page evaluates images vs how they are evaluated in other parts of the website, and i don’t think that was ever resolved.
i think if you want to get the best computer vision results during upload, you need to make sure things fall within the center square of your image.
elsewhere in the website, you can get better recommendations after upload because it evaluates a little differently, and (as has always been true) it will also narrow down the universe of possible suggestions based on the iconic taxon of the observation. so if you didn’t get any good suggestions during upload, go ahead and specify a high-level ID like plants or birds, and then check the suggestions again after upload. most likely they will be better.
…
on another note, i’m guessing that computer vision suggestions, especially for distant or blurry subjects will probably get worse and worse in the future, not because the model is getting worse, but because it has more species that it knows about, and more and more of those species will have similar features that make them hard to separate easily.
That’s interesting about the center square crop. I’ve been mostly cropping to square for a while now so my pics don’t look weird on the Explore/Identify page—good to know it has an added benefit!
I guess my main concern was not getting better or worse suggestions, but about the CV not showing any suggestions at all, which I don’t recall happening before. However, it may well be something that is not new and I’m just noticing it now for whatever reason. Or, to your other point, the CV now knows about so many more and similar species that it isn’t as confident about its suggestions as before.
that’s not new. even many years ago, if you were to try to get suggestions for cryptic species against a nondescript background, or a tiny, indistinct subject, or a mix of miscellaneous plants from a distance, etc. then there would be a higher than normal chance that you would get no suggestions because CV wasn’t confident enough to provide suggestions. i’m not sure what the threshold is exactly, but when computer vision suggestions for a taxon falls below a threshold, they won’t be suggested. so if all possible suggestions fall below the threshold, then you get no suggestions.
