A while back I posted a photo of a Common Hoopoe that was basically just a blurry smear. To my amazement, the iNaturalist machine ID recognized it without a hitch. I was just curious if other folks had ever had similar experiences with the machine ID suggestions. I’ve seen other threads about the machine IDs being horribly wrong, but what about amazingly right?
yeah, tell it to this observation
The fact any kind of program can do so well across so many forms of natural life is really, quite incredible. It does not do well on hard-to-ID groups such as spiders and moths, but for the majority of observations it can give you a satisfactory ID or something that will get it noticed by more knowledgeable folks that can correct it.
Actually I find it to be pretty good on moths, because often pattern recognition is what is needed, and it is good on that.
There is no computer vision ID tag on that, and the top computer vision suggestions are puffball fungi (makes sense given the image). Looks more like a joke, test, or accident by the observer.
i can’t think of a specific amazing ID off the top of my head, but i am generally impressed by the computer vision suggestions for visually distinctive birds, insects, and plants.
it’s not exactly an accurate ID, but i was impressed that my observation of a bee next to a hummingbird feeder (https://www.inaturalist.org/observations/62374363) was suggested by the computer vision to be a hummingbird. it was wrong, but it was still right in a way.
I think it looks more like a sea otter.
I am always amazed by the surprisingly good suggestion from the AI. Yes, I’ve had a couple that were hilarious, but those are by far the minority.
It can be a bug, there’re lots of ids 100% based on machine suggestions and yet not labelled as such, as I remember there’s a bug report that app doesn’t always report that? Because those are always phone pics.
I was pretty amazed that the CV got this one of mine of a Nessus sphinx moth. I knew the species in person, but the photo is a just splat of motion blur! I wonder if the model is trained on blurry photos of sphinx moths darting around.
Wow, I can’t even tell that that’s a moth, lol.
It Id’ed a Texas Brown Tarantula out of this cropped in image. https://www.inaturalist.org/observations/62610237 First on the list of automatic suggestions. I’ve never seen it be this accurate with such low quality.
Meanwhile I can see what the Computer vision had to pick from this photo https://www.inaturalist.org/observations/19358519 to get it is Oenothera, but I remember clearly how amazed I was to learn from the top suggestions that it is Oenothera before it was IDed by the community. It was my first day on iNat, and the plant has puzzled me already for years. I labelled it for myself strange exotic seeded miniature cucumber until then. ^^
I’ve seen other observations where the computer vision identified gray rocks as elephant seals. One time the rocks were gray, squarish concrete blocks. I view this is a relatively benign error, but the computer does make it.
Personally, the number of spot on and nearly right identifications by the AI is really scary stuff. I predict that within 5-10 years all the IDs will be AI, and human IDs will be considered too inaccurate to allow to interfere with the process.
It is not just the number of times the AI gets it correct, but also the huge breadth of its doing so, and the sheer speed at which it does it. And the irony is that it is all by giss (general impression of size and shape): there is no prior intelligence: keys, profiles, diagnostics, descriptions and significant features.
Knock it and criticise it while you can - Big Brother is growing fast!
It’s nice to be amazed an all, but one should keep in mind that the AI is just dumb pattern recognition. So if everybody takes blurry hoopoe-in-flight pictures, then the AI will learn to recognize blurry pictures of dark objects against a uniform background as hoopoes. It doesn’t use things like field marks, or likely location, or behavior like we humans have to. Since it uses photos that humans generate, however, it learns more about how we behave, than about what it ostensibly recognizes.
Sure. For now. It’s only going to get better over time…
to a non-birder, I assure you, those all look exactly the same to me… I couldn’t even tell it was a raptor in the first place. AI is already doing better than I am, better than many casual users will.
Yes, of course. But it recognizes pictures that are similar to other pictures that other people have taken before. It does not recognize an organism using the same means that we use.