Snails are exceptionally difficult to diagnose. Over the last week I’ve seen several misidentified (AI identified?) submissions of taxa I closely monitor in the Helminthoglyptini snail tribe. Two of the photos were of rocks. No sign of a snail or shell. The others were of taxa that are similar in appearance but were several hundred km out of range, and even from other continents. People rubber stamp Agree and turn them into research grade observations. Yikes.
I didn’t notice whether a taxon is AI identified on submission. I know this is community driven but there is room for improved curation.
What’s the question? This isn’t a bug in any case. The CV is far from perfect; the problem is people uncritically accepting it (if that’s even the issue here; I have found many cases where it actually gets it right, or at least close, but people somehow pull up wildly incorrect IDs). There is also the problem of people accepting what it says because it looks close, but it only works on taxa with a sufficient number of observations. So once one is in the system, more and more get IDed with the same name even though other similar species may co-occur.
We can’t forget that there are more than just scientists on this. There are citizen scientists, which also includes kids who might be learning about iNaturalist. In this day and age, theyre already addicted to the phone, so may as well have them be doing something (sort of) useful. amiright?
I can see snails being something kids are interested in, and maybe they are still learning the differences between rocks and living things because snails have rocks on them (who knows!). If its bothering you just ignore it. It will remain un-ID’d and you can breath easy.
I came across one of those among my bookmarked URLs. Skimmed all their other obs, it was only this one, which was optimistic about a snail among the rocks. The others all had a clear focus on identifiable animals. Anyone can make A Mistake.
The great thing about iNat? You can be part of the solution! If things are misidentified, you can help correct them. Regardless of the cause of the initial wrong ID. And if the cause was the CV model, improving the data it learns from helps minimise future problems. (Though yes, on some things it will always struggle. Show it a pile of rocks and it’s required to come up with some answer because that’s how it works.)
For one of these problem obs – the Phoenix snail was only offered if NOT Nearby was chosen. And then only 4% certainty, which I find a reasonable answer from CV – among the rocks was a little lump which could have been A Snail. For Nearby the first choice was an otter (for a river habitat) but also only 4% certainty.
Responsibility for the chosen ID rests on the iNatter whose name is attached to the choice, not on CV which is a tool. Blame the tool for your / their own mistake ?
Geomodel hexagons are 60 kms across – which means my Cape Peninsula is inside one hexagon, with a small corner overflowing into the next hexagon. So I use it thoughtfully, local Endemics give it hiccups.
Moderator Note: I’ve moved this from Bug Reports to General, as the Bug Report template wasn’t filled out and this doesn’t seem to meet the definition of a bug:
“something is not working the way it is intended to”
I will openly admit that if I don’t know the group well, AI is “pretty sure” with a genus and the thumbnail looks like the same kind of animal, I will just accept and move on. Always only at genus level (so that pure agreement doesn’t bring it to RG). I hope eventually someone will look at it and correct it. There are many IDers who filter for only specific taxa, so giving only very broad ID may mean that nobody ever sees it.
If you post a bad photo looking up into the leaves of a tree, the Computer Vision will often suggest a bird. This is because a lot of bird photos are like that.
“something is not working the way it is intended to”
This is a procedural problem/bug. I am providing feedback in the hopes it can be improved.
The AI being overly optimistic/wrong is only one of the problems. - Limiting the likelihood cap could help. More training on identifying rocks might help too. It was announced early on that there would be errors. I just wish the critters I work with weren’t in it /grin
In the app I got confused about identifying. I suspect I’m not alone. There isn’t clear way to just type in the id. The app drove me through the “Identify” option and AI popped up whether I liked it or not, and if my critter of choice wasn’t found, it wasn’t clear what to do next because I couldn’t just type in the name. If I were someone less savvy, I may have just picked Unknown, or the first thing on the list to complete the observation out of satisfaction that it was done with in the app.
If they were using the app, it is very easy to select a bunch of photos, including ones that were not intended. A couple things happen. The user could miss they have the ability to delete a photo from the observation(s), or, they grouped to the wrong observation. I’m betting the Observation began this way.
If you have suggested improvements to the app it will be helpful to everyone to use the feature request template which asks some specific questions to help everyone get on the same page. Screenshots and mock-ups are super useful. For example, are you referring to the Android, iOS Classic, or new iOS app? on Android there is a “Species Search” bar and in the new iOS app, there is a search icon (magnifying glass) in the top right of the suggested IDs page. If the new iOS app, are you in simple/default or advanced mode? (I’m not necessarily asking you to answer the questions here, just using it as an example and letting you know how to move forward with specific requests or recommended changes to the apps or website.)
The CV isn’t trained on rocks because iNat’s categories only include living organisms (or objects made by living organisms).
As noted above, the CV does not have the option of saying “there is no organism in this photo”. It has to suggest something. If you give it a uniform block of color to identify, it will still try to find a match in its training set. It doesn’t know what an organism is – and thus how to tell whether there is no organism. It is an image matching algorithm only.
Now, I suppose theoretically it would be possible to train it on observations marked as “no evidence of organism” in order to teach it to recognize images which contain no recognizable organism. However, it is likely that this would create some problems, because there are a lot of images in which non-living matter takes up the majority of the image and the organism is only a tiny spot somewhere in it. Teaching the CV to distinguish between images that have meaningful tiny blobs and those that do not would probably be quite challenging if not impossible.
Another, somewhat more feasible option might be to change the suggestion mechanism to have a confidence threshold below which it would not make any suggestions at all. I suspect, however, that this would result in both too many false negatives and too many false positives to offer a meaningful improvement in the quality of the CV IDs that people apply to observations.
I noted some of the challenges above – it doesn’t at present “know” anything about the taxa it suggests. It doesn’t even know what objects are. All it knows is that certain patterns of color are associated with certain labels, and it compares the images it is presented to see what patterns of color are the closest match to the images it has been trained on.
If you could combine it with the ability to crop, I can see it could be useful: “We can’t spot anything recognisable here. If the organism is only taking up a small amount of the photo, please use the cropping tools to focus on the relevant part, then try again.”
And if it could pick blurry photos and encourage people to retake - that would be wonderful! Albeit unhelpful for those uploading later…
I’m not a Seek user, but I have some vague recollection that it might have something like this – that if it can’t find any good matches it encourages users to take another photo?
It was in the back of my mind while writing that a low confidence score from the CV could be used not just to limit suggestions but embed some nudges to users to help them make better observations (or at least upload photos that are cropped and in focus). Though I don’t know if this might have an unwanted effect of reinforcing the belief that the CV is intelligent rather than a tool to be used critically…
I mean this would actually work. If you include a “nothing” taxon in training and give it a sample of photos of various habitats with no life in them - which is not actually as steep of a task as it looks like, because it’s basically just sand, water, snow and rocks, everything else has something obviously living in it.
The question is whether it would be worth spending the effort, considering that it would only help … a very specific subset of observers, who are probably not likely to stay here for long.
There are some taxa which require geosense, as well as having characters which are devilishly hard to distinguish from other taxa. That’s a tough thing to balance. I wonder how well the CV distinguishes monarchs from mimics.
What are you insinuating? I don’t think this has a place here.
Perhaps people who maintain the UI will recognize the value of something and decide whether it warrants their time. I don’t have a specific change, how could I? What I’m experiencing are usability issues more than anything. My expectations were different than reality. I feel like I’m being hushed, so I’ll bugger off.
What we are pointing out is that it’s going to be easier for the UI people to “recognize the value of something and decide whether it warrants their time” if they understand not just what isn’t working for you, but how you expect it to work. They aren’t mind readers and people interact with the UI in different ways, so what is obvious to one person may not be obvious to someone else.
There also isn’t just one interface – there are at present three apps and a website, all of which handle the ID prompting process slightly differently, so nobody will be able to address the difficulties you are encountering if they don’t have a bit more information.
What we are also pointing out is that some of the things you mention are known and are likely not quite as simple to fix as they might seem.