Seek - Should I submit photos to help the algorithm?

There’s clearly a discriminator in the (I assume) neural net that trains the IDs in Seek. Is it helpful for the devs to have false positives posted? (see this elephant/dog and a chair):

The AI isn’t made to recognized humans or human made things, as they come in all shapes and sizes. Because of that, the AI couldn’t recognize the shape of the chair, and since the dog is only a silhouette, it all blended together. I’m not sure if there’s much that can be fixed here.

Note for Forum Mods: Possibly move this to #bug-reports

Hi @buriedanimal - observations made with the Seek app aren’t used for computer vision training unless you choose post them to iNaturalist as well. There are over 45,000 photos of dogs (or evidence of dogs) on iNaturalist for it to randomly select from for training already, so it’s kind of funny how wrong it can still be. It’s up to you whether you are interested in posting that to your iNaturalist account. Perhaps the training set needs more lower quality / unclear focus / low light photos to help with these types of situations.


i doubt that this particular photo would be particularly helpful for iNaturalist’s computer vision training. the main reason is that iNaturalist isn’t really geared towards observations of captive organisms. the other thing is that there’s a lot going on in this photo. (there are multiple subjects, your intended subject is less in focus than other objects, your intended subject is relatively small, etc.)

i think the elephant in the room that Seek is pointing out here is just that any computer vision implementation is probably going to have a hard time trying to figure out exactly what you are wanting it to see in certain situations.

just for comparison, Microsoft AI for Earth’s species classification model also has a hard time with certain kinds of photos that iNaturalist’s computer vision has a hard time with.


So it’s okay to post there, if only to help with training? I don’t want it to be mistaken for fake spotting.

The pun did not go unappreciated :elephant:! Depending on what kind of CNN is being used, false positives are a layer that needs to be tested against - if it’s too dark or cluttered, the hypothetical goal is to never ID it (as an elephant or anything else). I just didn’t know if this would contribute to the model or not.

1 Like

here’s a description of how the last computer vision model was trained:

it doesn’t sound like they do anything like this-is-not-an-elephant training. (where would such a training set even come from that sort of thing, unless you just assume that any given photo of, say, a dog absolutely will not also contain an elephant?)

i’m not sure how they do their positive and negative testing.

i’m not sure exactly what you mean here, but if you posted your observation from Seek, you would want to identify it in iNaturalist as a dog, not an elephant.

1 Like

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.