Step 3: create & resolve bug if the inat folks are committed to their steadfast reply that this is not a bug and is infact intended behaviour of the cv instead.
I cannot attempt to reproduce the bug because you added a new ID.
also, thereâs no need to be rude.
hallucination is a feature of generative AI. CV is built for pattern recognition; it has lots of false positives but doesnât invent new ideas.
however, hereâs my best guess for what happened:
CV by default tries to recognize ANY taxon. But if the observation already has an ID it will narrow its results to descendants/relatives of the existing CID. as the ID was of an insect it suggested insects. so, user error followed by CV failure to explain itself to user
Is writing, in words, how the inat community views these issues (not a bug) a rude thing to do? I didnât know that telling the truth is somehow offensive
Sorry but the fault here lies with the observer accepting a CV suggestion for a butterfly - when they were interested in the flower. Even starting with a broad ID - is a flower = angiosperm, would have started them in the right Tribe. And their obs would no longer be trapped in Life.
We will always get a better (even good) suggestion from CV if we start from taxon and location.
Plenty of obs trapped at Kingdom disagreement due to âblind acceptanceâ of the wrong homonym. Currently haunted by Aizoon which can be Genus (what I and my location expect) or Subgenus which is a ârandomâ and totally different plant in the other / Northern hemisphere.
Human error. And - to be fair - while the problem obs is at Life, CV is still offering me the butterfly.
As astra said - CV is pattern recognition. Pixels. CV does not âsee a butterflyâ or âsee a flowerâ. That is our human interpretation of - what do I see in this picture ? On behalf of CV this view of the flower is disconcerting - we are looking into the flowerâs face, I can see the dangly bit to the left. But, and it is a big butt. There are leaves blocking the view of the centre of the flower. If you narrow your eyes and step back ⌠you could see ⌠2 (almost symmetrical) red wings and a green body âŚ
Soon after we released the computer vision model, Cal Academy asked us to demonstrate it at a table for their Nightlife event one week. So I grabbed a few leaves and lichens from Golden Gate Park to put on our table. One was a dried up lace lichen that had curled into a ball. The CV would consistently ID it as a hummingbird, which confused me until I realized it had been trained on some hummingbird nest photos, which are cup-like structures that often incorporate lichen.
Once I straighted out the lichen, the CV gave me the correct ID.
When I uploaded this plant in front of a waterfall I remember the cv suggested among others Blue Whale and Andean Condor. I canât reproduce this today.
Thatâs how I feel every time the CV âlooksâ at a photo of one of the five species of black-and-white Skimmer dragonflies around here, and tells me that itâs an Atlantic Surf Clam.
@tiwane, are you sure that you guys didnât hire John Cleese or Eric Idle for a few days somewhere along the line?
(EDIT to add: that hasnât happened for the past couple of CV model releases, but I still remember it fondly. Mostly.)
