Computer vision questions

I am doing something new with my inaturalist observations. For species that are not yet covered by computer vision due to too few observations, I do large numbers of observations in the hopes that I can get the species over the threshhold. The problem is I don’t really know the threshhold. It used to be “20 research grade observations by 20 individuals”, but was more recently changed to “100 photos of which at least 50 are research grade” (not an exact quote). The previous criteria seemed clear, but the new criteria explantion seems likely to be imprecise. If taken literally, it seems to indicate that someone could take 100 photos of a single specimen and if that observation is research grade, it’s accepted for computer vision. I can’t believe this is true. Also that statement seems to indicate that non-research grade observations are used for training, which seems really bad, but I suspect that isn’t true.

Can someone provide a very clear explanation of the criteria used for inclusion of a species in computer vision?

Also, I’ve noticed that some species that should be covered according to the most conservative interpretation of the new criteria aren’t. The distinctive kermes scale moth had 100 research grade observations 2.5 months ago (184 now), but still isn’t covered. How often is training done? Why isn’t training done immediately after the threshhold is met?

1 Like

what-image-s-are-used-for-training-computer-vision/3307
https://forum.inaturalist.org/t/what-image-s-are-used-for-training-computer-vision/3307

https://zpascal.net/cvpr2018/Van_Horn_The_INaturalist_Species_CVPR_2018_paper.pdf

kermes scale moth https://www.inaturalist.org/observations?place_id=any&q=Euclemensia%20bassettella&subview=grid

https://www.inaturalist.org/pages/computer_vision_demo

I cant answer this one precisely other than to say it is an infrequent process as it is extremely computer intensive to run the training model.

1 Like

maybe the staff can help with this?

According to the computer vision section of the FAQ;

Taxa included in the training set must have at least 100 photos, at least 50 of which must have a community ID.

I’m guessing that means that the observation has to have a community ID of that species?
I don’t fully understand this, so maybe this should be explained in more detail on that page?

1 Like