Vision Suggestion Updates!
We just updated our computer vision-assisted species suggestions! It's been almost six months since we quietly launched the first version, and despite not making much of a hullabaloo about it, you folks definitely noticed, as did the media. One of our favorite parts about our approach to teaching computers to recognize species in images is that our system is constantly learning from the iNat community. If someone chooses a suggestion that ends up being wrong, that turns into new training data, and if someone observes something the system doesn't know about, that turns into new training data too. In June, we released a system trained on iNat photos added up until May 2017. It could identify 17,246 species, and it had the right species in the top ten results ~78% of the time. The system we just released yesterday trained on data through August 2017, and while it's only slightly more accurate (right answer in the top ten ~81% of the time), it knows about 20,217 species, so that's a 2,971 species improvement!
To give you and idea of what's changed, here are some of the most-observed new species the system can recognize:
We were very happy to see a bunch of species from outside our core areas of the North America and New Zealand in there, and we were particularly impressed with Sphenomorphus indicus, a lizard that has experienced a surge in observations this year thanks for iNat folks in Taiwan. A lot of species have years of observations, but only just passed our threshold of having Research Grade observations by ten different people, but that lizard really just became super popular this year. Go lizard.
We also made a slight change to how we use nearby observation data to add suggestions and sort them: we reduced the radius of the search, so hopefully it will make better "nearby" suggestions. It is not excluding suggestions that have not been observed nearby, but that's certainly something we're considering since so many of you have asked for this.
Anyway, a big thanks to all of you for making all of this possible. We couldn't provide this kind of service without all of your hard work making observations and adding identifications.
Comments
I am very interested in this development. I am hoping to be able to enable the unskilled local Tamborine Mountain resident person to identify our local plants and animals by them taking one or more pictures on their smart phone via I naturalist and the application returns the name and common name. Earthtracker.net is my now defunct/ superseded website.Click on "FINDER" menu item.
As usual - great update! This provides more encouragement to folks to participate in identifications and verifications.
Also, a pretty interesting discussion on the app: https://player.fm/series/naturalistics-podcast/naturalistics-008-inaturalist
I did mention to those guys about the importance of the website and community (this was left out of the podcast for the most part)... Nonetheless, they discuss some interesting topics!
Very neat!
I have to say I LOVE the ID suggestion feature. It's not perfect, and tends to suggest endemic NZ species for my Texan observations, but it is amazing. I volunteer with a nature-based after-school program with 5th and 6th graders, and the app has instantly identified things they've found. For the caterpillars, I was able to show them pictures of exactly the kind of moth they turned into, within minutes of them discovering the caterpillars.
Nowadays when I'm evangelizing iNaturalist to random strangers, I'll mention the ID feature and how great it is, in addition to the million other benefits. Great work on that, guys!
I'm impressed by how well the suggestion feature works. It correctly IDs species even when I post some pretty poor photos. It's great to hear that the number of species that can be identified in this way is expanding.
Great update as always. Any data to share on accuracy presented by major taxon group (insects, birds) or geography (North America, Africa, etc)?
@nanofishology yeah i think there should be a hard 'no' on showing things that far out of range, at least in places with lots of iNat observations (North America, New Zealand). But we are getting there. And yeah it's pretty amazing.
What is the best way we can help the IDs improve? I have been waiting for the recognition feature even when I know what the specimen is, on the grounds that clicking on the one I know it is helps with the training. Is this true or not? I am quite impressed with how it does with plants not blooming. It's fun to see what it will come up with.
I don't think choosing from the ID algorithm helps the algorithm any better than choosing ID on your own. The devs can correct if I am wrong but my understanding is the best way to help the algorithm (and everyone else too) is just to add lots of observations with good photos, and to help others identify their observations and get research grade.
Can do, thanks!
@charlie I'm also noticing my African observations for arthropods (and even some birds!) frequently come up with taxa endemic to North America. When I got the New Zealand suggestions, I wasn't sure if the algorithm responded to user history, since I do have some observations from New Zealand, or if there is just so much data for the New Zealand species that it bumps up those suggestions.
This is so awesome. It makes learning much faster for me as a user because I can quickly learn what other things to ID against instead of just the 1 organism that I hit in a field guide.
@nanofishology, the algorithm currently doesn't distinguish by location at all except noting if something was observed nearby, but there is hopefully going to be a future update that adds that functionality somehow, because it is a known problem (and the blog post mentions it)
Very interesting! @charlie Do you know what its range for "nearby" is? 5 miles? 50 miles?
I don't. It used to be very wide, I think at least 100 miles, but sounds like it shrunk now
Add a Comment