New Computer Vision Model (v2.16) with over 1,000 new species!
We've released v2.16 today with 1,389 new species (94,758 taxa up from 93,369). This new model was trained on data exported on September 1st, 2024.
Here's a graph of the model's release schedule since early 2022 (segments extend from data export date to model release date) and how the number of species included in each model has increased over time.
The graph below shows model accuracy estimates using 1,000 random Research Grade observations in each group not seen during training time. The paired bars below compare average accuracy of model 2.15 with the new model 2.16. Each bar shows the accuracy from Computer Vision alone (dark green) and Computer Vision + Geo (green). Overall the average accuracy of 2.16 is 90.7% (statistically the same as 2.15 at 90.6% - as described here we probably expect ~2% variance all other things being equal among experiments).
Here is a sample of new species added to v2.16:
Comments
Wow, congratulations! Interesting that quinoa was added only now; I didn't know there were that few observations of it.
Nice! Excited to see one of the species I photographed got included; Musculium securis, a tiny little freshwater bivalve. It's always fun to browse the "other animals" section too!
I was always love browsing the new additions to look for ones for which I've contributed photos. Fun!
I'm excited to see Gafrarium tumidum in there... Less than two months ago I noticed that it had been resurrected from synonymy with G. pectinatum (in June 2022) and asked to have the taxon added to iNat. Despite it having an overlapping range (so there was no atlasing the split), there are now enough RG observations to have the species included in the CV!
Extremely exciting.
There are 490,865 total species ever observed on iNaturalist and 405,155 of those are research grade.
When you say "94,758 taxa up from 93,369", do you include genera, families, orders, etc. Or does this mean 94k species?
its distinct taxa - so could include a genus etc if nothing downstream was in the model
Got it. So basically any leaf node
It's really interesting that the increase is still linear - I would expect the number of taxa added to start to tail off at some point as the "easy" taxa have been added, but taxa are still being added at about the same rate even two years after the transition to regular model releases. Very cool (and helpful)!
@guerrichache
I was about to say the same thing! It’s gone up from 29 observations in May to more than 100
https://www.inaturalist.org/observations/217272808
'on the Peninsula it seems to now be restricted to this small locality' - comments John Manning.
I expected it in the 'next' CV update. Thank you!
Wow!! Congrats iNat!!
woohoo, a few hoppers more
Relaying this question from another person.
Are the model weights held proprietarily or can the model be downloaded somewhere?
@erquint. Not currently, but we would like the capacity to do that. The data archives iNat makes available under CC-BY-NC licenses (e.g. the GBIF archive) exclude about 30% of data that users have chosen not to share. The models we currently train are trained on all data (including this 30%) so we can't license or share them. We are woking to increase our model training capacity (bottlenecked by hardware/$ at the moment) to train an alternative set of models that are only trained on licensed data and can be similarly shared,
@loarie, quite understandable.
Appreciate your response very much.
Add a Comment