Description of need: Often the Computer Vision will make suggestions that aren’t actually what the subject is. There is a bit of a bias toward animals, so a common issues is that you’ve taken a photo of a tree or plant, and something else in the frame (recently a lightbulb in one image) triggers the CV to suggest an animal instead. Or someone has taken a photo of a plant that happens to have an animal and the CV tries to identify the animal instead of the plant.
It would nice to be able to tell the CV to only suggest plants or only suggest animals or only fungi, etc Keeping it to read categories only.
Feature request details: As above, when the CV system makes suggestions have an option that says something like limit CV, and from there a very short menu that sticks to broad categories (like kingdom level), realistically I’d suggest plant, animal, fungi, and other as the options.
When selected the CV system would redo its search, but limit results only to potential identifications that fall within that category.
for what it’s worth, i’m fairly certain that vision suggestions will be limited to only taxa within the “iconic taxon” of the observation, based on existing IDs. so you can use this behavior to limit suggestions, although it may not work in some situations such as when the existing IDs for the observation are the wrong “iconic taxon”.
This function already exists in the (android) app: you can filter suggestions for whatever taxon level you choose. On the website, one way to go about this is clicking on the ‘compare’ button of prvious ID’s and changing the filter options, or if there is no iconic taxon/other ID’s yet you could first suggest the order level as an ID and then click ‘compare’ on your own ID.
I’ve asked what the impact is from the introduction of the geo_model in the other parameters that effect the results of the computer vision model, but did not get an answer (yet). I expect that the results will be limited to the icon taxa of that observation, as stated in the answers above. https://forum.inaturalist.org/t/log-computer-vision-taxon-guesses/44224/7
We discussed this in depth prior to the feature request being approved. It’s unreliable, doesn’t work in many cases, and requires a double ID, which may users don’t or won’t do.
Having an option at the outset to limit CV suggestions to a particular broad category (eg, plant, animal, fungus) would have many benefits.
This approach (entering an ID to get an iconic taxon-restricted CV suggestion) has always worked for me, though I only use the web. If there is a bug, it would be useful to have a reproducible example, because currently the system is designed to allow for this functionality.
Whether the way to get taxon-restricted suggestions could be improved is a valid question I think but is a fundamentally different issue than a bug in the existing feature.
So I tried the two-step approach recommended here, identifying it first as an insect. Even now that I have annotated the life stage as juvenile, the CV is suggesting insects that do not have a larval form resembling the organism in the photo.
You have a pretty restrictive place filter set it looks like. For organisms that are less likely to be observed, this can knock out many potential options. I’d try setting it to North America or similar since that would give some plausible options. The suggested approach seems to have worked to accomplish the goal of this request in your case at least (refining the CV responses to an iconic taxon and ignoring moss).
You are using “observations” as a source and getting the most common insects in those place filter. You should try “visually similar” instead. But as this is becoming about how to use the Identify model, we should make a separate thread if any further conversation is needed so as not to get this feature request off topic.
See similar previous conversation here for more info:
Okay, thanks. I didn’t realize from the description in the thread I replied to (sure, move this somewhere else if my concern is off topic for it) that the default was anything other than the CV.
“Visually similar” is at least getting warmer with insects and with butterflies and moths. It would be nice to be shown the larval forms of the suggested taxa, though I think that’s been discussed somewhere else already.
Anyway, I understand now to how to perform the basic operation, thank you.
Incidentally, among the Source options, along with Observations, RG Observations, Captive Observations, and Visually Similar there’s Checklist. It’d be nice if hovering over the option could reveal an explanation of what it means.
