Add a "search similar observations" function that draws on computer vision similarity rankings

Platform(s): website

URLs (aka web addresses) of any pages, if relevant:

I found these threads, which describe something adjacent to this request, but I think they are fundamentally different:
https://forum.inaturalist.org/t/reverse-search-images/3004
https://forum.inaturalist.org/t/inaturalist-computer-vision-ranking-and-combined-observations/26405/10

Description of need:

The purpose of this would be to assist in finding other existing observations to identify/organize, rather than assist in identifying a single observation (which I think is what some of the past feature requests are about).

Here is an example of why I am requesting this.

This is an observation of a very unusual cockroach: https://www.inaturalist.org/observations/194235569

I want to be able to find other possible observations of similar looking cockroaches on iNat. I cannot perform this with a text search because any observation of a similar looking specimen would be misidentified, or so broadly identified (e.g., Blattodea) that it would take very very long to find. Using reverse image search (e.g., on Google Images) is basically useless for this because I only get suggestions of beetles or totally unrelated cockroaches.

I think this feature would be useful because it would massively assist with identification or sorting observations into projects.

Here’s another example:

I made this project to collect all curated Lamproblattidae observations in one spot: https://www.inaturalist.org/projects/lamproblattidae

I did this because Lamproblattidae were poorly identified on INat previously (despite being moderately common and large insects). So i searched through every neotropical observation of Blattoidea cockroaches identified between the super family and family level. This took me multiple hours a day for about 2 weeks to go through all the results. Although the computer vision suggestions would only be slightly better (it probably would have only filtered out a lot of the Blaberoidea and Blattinae observations erroneously IDd as Blattoidea) it still would have saved a lot of time I think.

Feature request details:
Using the first example above, it would be great to have a feature on the observation page (e.g., https://www.inaturalist.org/observations/194235569) where there was a link that says “search similar observations” and it simply links to a list of similar observations from the computer vision model, ranked by similarity. I suppose this differs from the current implementation of computer vision suggestions in that it would suggest similar observations, as opposed to similar taxa. This is an important distinction because I can already search for the taxon name (Oulopterygidae) and I do not get similar observations (because of the ID problems mentioned above).

Without technical knowledge of how the AI works I cannot know if this would be a huge increase in the functionality of the model (and thus impractical) or if it would be a new implementation of the existing functionality. My understanding is that the CV model ranks the images from observations, but maybe only observations from existing identified taxa (??). If that’s true then I think what I am requesting would be a big expansion because it would require one photos from all observations to be part of the model.

There are Phylogenetic projects
https://www.inaturalist.org/journal/jeanphilippeb/73398
but not for Blattodea

1 Like

in this case, what you need is an observation-level or photo-level visual search, not a search that draws on computer vision similarity rankings. this is because i don’t think computer vision will have any concept of that family or its subtaxa because there are not enough images in the system for these taxa to make it into computer vision training.

if you wanted to train your own computer vision model, that’s something you can always do on your own, and then you can have your model screen roach images in iNaturalist for you.

1 Like