@pisum In my problem (appears below after being moved here) I am getting exactly the same list of things as you got for your blank image. So this looks like CV is getting a blank image instead of the real one? I however wonder why for the microscope things, it does this and for other pictures it just gives up? But the photos do come from a different camera.
My laptop is 100% AMD, so I guess that eliminates the nVidia theory.
The most important thing to learn is that it’s a Firefox issue - works fine in Edge, so I can just use that, all browsers are the same anyway.
On several observations from microscope, CV gave me nonsensical recommendations - and those are exactly the same list for each photo. The same list also returns from photos that it previously made sensible guesses for.
I tested CV on simple photos of birds (that it previously had no issues with) and it tells me “We’re not confident enough to make a recommendation.” Haven’t found any photo on which it would give me sensible guess, so it’s completely useless for me now.
that’s interesting. if that’s the case, then maybe you should send one of the microscope images to iNat staff. maybe it is something strange about how firefox handles certain images after all.
or if it’s failing for all photos now, you could check for a setting in Firefox related to canvas fingerprinting. if you turn on the feature to resist canvas fingerprinting, that should cause these sorts of problems on any image loaded in that upload page (since that page sends stuff to be evaluated by the CV after some manipulation using the canvas element in the browser).
Adding iNaturalist as an exception to "advanced tracking protection” solves the problem. Not sure what is Firefox trying to protect me against without me even knowing about it, but that’s the current trend in everything. Thanks, this is now solved for me.
my understanding is that the way your machine renders a particular image or text on a hidden canvas element is unique to your particular hardware and software setup. so folks who want to track you can tell your browser generate such an image and turn that into a hash (fingerprint). each time this occurs, the same hash will be generated (unless you change your hardware / software), allowing it to be used to track you across the internet (regardless of whether you use private/incognito mode), separate from tracking cookies and other such standard trackers.
fingerprint protection blurs the canvas, injects noise, or otherwise tweaks the canvas to prevent the resulting hash / fingerprint from being unique / consistent, thereby preventing tracking.
Wow, I can’t even imagine how this works. It’s interesting, I used to be a lot into programming 25 years ago and new all sorts of hacks, but since then, the world completely changed and most of my understanding is now entirely useless!
Screenshots of what you are seeing: The same list of suggestions was given for the middle photo, but I could not keep the suggestion bar to stay down when I tried screenshotting for some reason. Point being, all 3 photos gave the same list of clearly incorrect suggestions.
Description of problem:
Step 1: Similar issue as previously reported ( https://forum.inaturalist.org/t/excessively-generic-really-off-vision-suggestions-on-website/14637 ). Some images give wildly incorrect, identical CV suggestions. This did not occur with all images, but overall, the accuracy of the CV seems to be degraded. Photos that clearly depict a species I know the CV has identified reliably in the past from equal or lesser quality photos do not have the correct ID confidently (e.g. the first) suggested.
Update: The issue has been replicated with the last 5 observations I’ve uploaded. Always the same or very similar list of suggestions for photos it would have easily given the correct ID for previously.
Also of note, this issue DOES NOT appear to affect the CV suggestions provided when making an ID on a published observation; i.e. once the observation is posted the suggested IDs are with the expected accuracy for a photo it previously suggested highly unreasonable suggestions for when attempting to submit.
I had this happen to me today. Everything I was trying to upload was being IDed as Evening Bats by the CV (none of which where bats of any kind). I traced it down to a particular Firefox setting in about:config, privacy.resistFingerprinting = true
I am not sure exactly why this causes the CV to behave this way, but it is repeatable and there is no way to opt out of that setting per site so it means that it has to be disabled globally in any Firefox based browser (I am using Zen) to use the iNaturalist bulk upload screen which is frustrating.
Did that but back in full force today. Started last night with one or to way out suggestions, today everything is a worm or a bat again. Off to Edge (ugh) to re-upload 300+ images of wasted data costs. Wait two weeks and see if Firefox accepts the CV as it were…
based on your previous posts, your issue seemed to be related not to the fingerprinting issue, but possibly to Firefox issues with Nvidia GPUs. others seem to have said that restarting the browser could resolve the the issue in the short term, but long term fix would probably have to come from Mozilla developers (or possibly from your OS developers – if you’re using Windows, i suspect MS developers have done something recently to the way Defender interacts with memory that could be causing problems with all sorts of things).
Description of problem: On all 4 of my observations today – poison ivy leaf mite, a ruellia, a furrow bee, and crepe myrtle – CV keeps suggesting eastern red bat as the top choice. The correct suggestion is put forth after I’ve made an ID first, but not before.
