For the most part the CV works really well, but every time I’ve photographed deer the suggestions it gives are way off.
For example:
Could it be that large, mobile, city organisms are likely to have a number of different backgrounds that mess with the CV’s training?
Just out of curiosity, did you specify the location of your observation before asking for an ID by the CV?
I notice the missing “expected nearby” witht he “visually similar”. Maybe you did not do this? Usually when I ask for an ID by CV and forget to specify a location I get IDs that are way off.
Good catch, I’d forgotten.
After adding the location the first suggestion is now correct. The rest are still quite far off, but that matters less
Another issue which will affect CV’s suggestions is cropping. I notice that the deer in your example is a small animal image in a wider picture. Cropping much, much closer to an animal (or plant) will always improve CV suggestions by a huge degree. I’ve noticed this with everything from deer to birds to moths to daisies.
Totally agree. I’ve photo’d quite a few deer and never had issue with their IDs. Sharpening a blurry image can also help if you have that capability.
We uploaded this photo on iphone app and web just now without search nearby and first choices were still white tailed deer
If you don’t know which deer - start with the taxon level you are comfortable with.
Yesterday was a jellyfish obs. Zoomed in and thought for a bit. It was transparent - because jelly - but those ‘threads’ are tentacles, and X marks the spot is red channels in the blobby bit.
But, cleverly photographed against a turquoise (and patterned background) to force details to show - CV’s first suggestion was Homo sapiens for the turquoise. And frankly, mine too!
There was an example recently where an OP uploaded an image of a big sunflower with a tiny grasshopper off to one side on the ray flowers. iNat’s CV identified it as a Petrophila moth because it has been “trained” that, more often than not, images of a big sunflower with a small insect off to the side are showing that genus of moths. I tested a cropped version of the same image and it properly concluded it was an Acridid grasshopper. CV analyzes the whole context of an image as well as any details it can glean from the subject animal.
I suspect in the case of your deer upload, CV is attuned to the context that a spindly, 4-legged animal in a suburban setting is likely to be a deer, even if it can’t see the precise “field marks” of a White-tailed Deer.
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.
