Hi everyone!
Sometimes I see an observation where the user has opted out of community ID, but the user’s ID is incorrect. So there could be upwards of five or six people posting a different ID, but the user’s (incorrect) ID is shown. I can provide examples if need be.
My question is, does the computer vision train using these observations? And if so, is CV trained as if the observation is the user’s ID or the (would-be) community ID?
Thanks, someplant
It’s my understanding that community ID is used: https://www.inaturalist.org/blog/31806-a-new-vision-model
Thanks for the response!
It looks like this would mean that it’s still worth adding an ID to them, because it can help CV train.
If the community ID is already solid, it should be fine (an extra ID that doesn’t change the community ID shouldn’t affect the AI training), but it also doesn’t hurt to add an ID, especially if there is some confusion due to conflict between community ID and the OP.
Hopefully I’m not changing the subject too much, but I recently noticed an “opt out” observation because it was well outside of the range of other species observations. Even though the community ID is not species level the ID from the original poster seems to be respected for the map because they opted out of community. This seems like a bug.
Species page with map: https://www.inaturalist.org/taxa/53395-Eriophyllum-confertiflorum
Observation: https://www.inaturalist.org/observations/35355247
That blog article is pretty clear that a single ID is sufficient for an observation to be eligible for inclusion in the training set. The community taxon is not a determining factor.
I wasn’t addressing which observations are or are not included in the model, but responding to the OP’s second question: “And if so, is CV trained as if the observation is the user’s ID or the (would-be) community ID?” In this case, the post described a specific situation in which there are already multiple IDs on an observation and the community ID conflicts with the user’s ID.
For more info, there’s this older article (one of the original iNat CV overviews) that describes the dataset creation process pulling RG observations from GBIF which uses the community ID.
The blog article “A New Vision Model” does not mention observations for which the observer has opted out of community ID. Assuming such observations are eligible to be included in the training set, only the observer’s initial ID would be relevant since subsequent IDs do not alter the observation ID.
How observations, for which the observer has opted out of community ID, are treated in the Computer Vision training?
Observations where the observer has opted out of the community ID can get used in training the vision model, but we use the community ID, not the observer’s ID. For example, if I think my obs is a dog, but everyone else thinks it’s a cat, if we use my photo in training, the model will learn that the photo is of a cat.
Thanks for the answer!