I find that the suggested IDs on iNat are indeed “visually similar” most of the time (though frequently taxonomically incorrect). However, just this evening (10 July 2020, 10pm PDT) I’ve noticed that suggestions have become either (1) extremely, wildly, not at all visually similar; in fact, all are “seen nearby” rather than “visually similar” – as if these are totally random picks, or (2) a very short list (1 or 2 suggestions only) of semi-visually-plausible suggestions. One example: see screenshot of “suggestions” for photos of Eriogonum incanum (https://www.inaturalist.org/observations/51172309).
Is computer vision not feeling well?
Mac OS10.11, Firefox 78.0.2esr, same results exactly on Safari.
1 Like
the score that computer vision returns consists of a visual match component and a location component. the location component looks like it’s working normally, but the visual match component seems to be returning 0 for everything.
there was some work on the computer vision API recently: https://github.com/inaturalist/iNaturalistAPI/commit/987cd11ced8bc340efff6772c55f5ea08516293c. i wonder if that could have affected this? @pleary might be able to troubleshoot this quickly, if he’s still awake. or @kueda might be more likely to be awake in his time zone.
1 Like
Yeah, it’s not really working for me either. There might be something to this.
Thanks for the heads up. I rolled back today’s release and it seems to be working again, so clearly we have a bit of a bug in our recent code. We’ll take a look at it on Monday.
13 Likes
I figured this would get pointed out by someone. My US experience was that all suggestions were the same species (always WILDLY off) no matter what the photo. (I think Monarch might have been the top species but I can’t remember)
But, it seemed to be working properly when I uploaded a photo and used the computer vision. I uploaded a dozen or more photos using computer vision successfully (as a check on my own instincts and a shortcut to typing).
It just didn’t work on any observation made for a new suggestion (of my own or anyone elses). So, if I was trying to suggest an identification for someone else’s observation, the computer vision spit out the same list of wildly off species it had mistakenly been doing. But if I downloaded the photo of the observation and initiated an upload of that photo and used computer vision, it worked properly and gave me suggestions that were more fitting with the organism in the photo. I did that a few times to check my own instincts.
Thanks, all, for taking a look at this. As of this morning (11 July 2020, 9am PDT) the example I provided is returning suggestions all in the correct genus, with the correct species at the top of the list. That’s how it should work for a relatively easy-to-recognize plant.
2 Likes
Thanks for reporting this. It was really weird… and irritating. I use that drop down box to save typing strokes (and typing mistakes and misspellings).
2 Likes