Show HN: Identify mushrooms from pictures | Hacker News Reader

Show HN: Identify mushrooms from pictures

champignouf.com

37 pointspingou49 comments

I'm not going to tell you how to live your life, but maybe you should have some sort of disclaimer: "not responsible for accidentally identifying death caps as chantarelles."

I'm going to tell you how to live your life, please put up a disclaimer.

This is the first example I've ever seen of an image classifier that could kill someone if it makes a mistake.

You may also want to consider listing all percentages and classification types, instead of just the highest match (if you don't do this yet).

I pivoted my idea on an allergen identification system after hearing the same thing from people on HN.

https://news.ycombinator.com/item?id=14637584

I think it would be accidental death, but it's a nice sentiment.

(BTW, I think this is a great idea. Just, yeah, disclaimer.)

For the curious: https://ww2.kqed.org/science/2016/02/23/this-mushroom-starts... This happens, and it's bad.

I'm somewhat baffled this actually happens that often. As a kid, I often went mushroom foraging with my family.

The first thing they taught us kids was to be very wary of death caps, and most mushrooms with gills in general. If it has gills and you are unsure, just leave it. Never had any accidents with wrong mushrooms and had lots of fun, at least until a certain nuclear reactor blew up and we had to stop.

Which reactor was it?

Chernobyl blew up and made foraging in many parts of Europe a rather risky deal.

Definitely, as I said elsewhere in a comment, I will add more warnings. The disclaimer is already there when you get the results, but it doesn't hurt to put some more.

Will also give some picture examples with mushrooms that look alike but with one delicious and edible and another deadly.

And maybe link to an article of some sort that spells out what horrible things the deadly one does to you because people can be pigheaded, slow to listen, etc.

chanterelles look nothing like an amanita :\

Yeah, that was my thought too. But I do have something growing in my back yard (in the fall) that is fairly similar to yellow chanterelle - probably jack-o-lantern. The gills are wrong, but I'd imagine that software to identify mushrooms from photos would miss that subtlety.

Sure that example is extreme, but the deadly amanitas do look like puffballs when young.

I would use a similar app for something inconsequential, like bird voices. However, mushrooms are often identified to determine whether they can be eaten. The penalty for mistake may be death. So I think I'll pass.

For the record, I've uploaded a picture of an aardvark: """I think it's: Panaeolus acuminatus Or perhaps Psilocybe zapotecorum , Coprinopsis variegata , Hygrocybe singeri , or coprinellus sect micacei or something totally different!"""

I think it's trying too hard.

I uploaded a picture of Lactarius indigo, one of the safest mushrooms to eat because it pretty much doesn't look like any other type of mushroom. (The trade-off for this certainty is that it is pretty bland tasting.) It is very difficult for a human to misidentify it. The results are below:

I think it's: Laccaria amethystina

Or perhaps Russula cyanoxantha , Lactarius indigo , cortinarius iodes , or Cortinarius alboviolaceus

So, the right ID is in the list, but it's not the #1 guess. This could be useful for trained mushroom hunters who just want to narrow down the possibilities to check up on. I definitely would not trust an amateur with it though!

OP here. It knows more than 500 species, results range from terrible to ok, but I will work on improving that, and add more information about each mushroom (like species that look alike but are not edible). Nobody should use it to know if a mushroom is safe to eat anyway, probably I'll add even more warnings as it may not be obvious enough.

If you know which are unsafe, you could have a threshold confidence, say the algorithm classifies one as shiitake with 90% probability but some unsafe mushroom at >40% probability (or some percentage that makes sense), it could say, "looks like this is a shiitake, but careful it could be <unsafe mushroom>, here's some pictures and info about <unsafe mushroom>."

Yeah, really the chief use for an app like this would probably wind up as, how confident are you that this mushroom will not kill me which is a slightly different problem.

I'm working on something similar as a hobby project.

What approach did you take to ID mushrooms? Model -> mushroom? Or Model->feature->mushroom? What kind of model?

Where did you get your training dataset?

Definitely add a big loud warning on the home page that this isn't a reliable.

Maybe clarify you're identifying the kind of mushroom from a picture. I clicked through, wasn't quite sure what the purpose was, as the text made it sound like the output would be "Yup, that's a mushroom."

It identified a picture of Toad (yes, from Mario) as Hypsizygus ulmarius.

FWIW, it failed at identifying my winecaps. It was a pretty easy picture too, albeit there were multiple in the shot.

It identified my thumb as one of a few possible mushrooms...

I'm no mycologist, but for an accurate identification of mushrooms often spore prints are needed. Also the color and shape can vary wildly over a few days, which makes this more tricky than a typical visual classification problem.

Are you working with or have you considered working with the folk at mushroomobserver.org? They have an absolutely enormous dataset. Hundreds of thousands of cataloged observations!

Nice name ;-) For those not speaking familiar french, pignouf is something like dumb-ass (and champi is champignon, mushroom)

What happens when this mislabels a poisoneous mushroom as something harmless? Where lies the responsibility? Can this app be used safely?

Imo it's quite different than a printed guide, because the app tells you what you're looking at instead of forcing you to interpret that by yourself like the printed guide does.

Seems especially tricky for users that don't understand that there will be tons of misclassifications with this kind of tech.

Even a mushroom expert can't identify an arbitrary mushroom from one photo. The reason is: either the photo reveals the surface features of the cap or the gill/hymenium features but not both. Same goes for the cap/stipe and any important underground features, like a volva (important for identifying many poisonous mushrooms). So...

Did you use a CNN to identify the mushrooms? (I assume so). Did you run into problems deploying the CNN to production? How do you intend to update it given feedback from your users, and the photos they submit?

I did use a CNN, but didn't really run into problems for that. I do not intend to use user submitted pictures, as I would have no idea myself what they truly are, if it's what you meant.

So I did a Google image search for "poisonous mushrooms" and it gave me a picture of "Galerina marginata" from Encyclopeadia Britannica[0].

According to the app: I think it's: Psilocybe zapotecorum Or perhaps Psilocybe subtropicalis , Psathyrella piluliformis , Psathyrella longistriata , or Kuehneromyces lignicola or something totally different!

Warning!!! Many mushrooms look alike and many are toxic! Champignouf makes a lot of mistakes! Do not eat mushrooms without being sure they are edible, you could die!

Thanks, but not helpful.

[0] https://www.britannica.com/list/7-of-the-worlds-most-poisono...

For the landing page I'd address the images used to link to App Store and Google Play. You're using images that are 2000 by 600px wide, but resized down to 500 x 50px... and the proportions are off. Should be an easy fix.

Thank you for your comment, yes I know I did it in a very dirty way, I don't really care about the landing page actually, more about the apps themselves as I'm a mobile developer.

I'd be curious to see how much traffic can a landing page bring though.

Be nice if you showed the uploaded picture and displayed confidence.

I uploaded a picture I took of an Amanita flavoconia, and it correctly identified it as the #1 choice. The mushroom wasn't particularly well visually isolated in the photo either, it had sprung up among a bunch of moss and twigs and only occupied about a quarter of the frame, so I found this fairly impressive.

This would be great if paired with a portable DNA sequencing device like those from Oxford nanowire: https://nanoporetech.com/products/smidgion

Is there a catalogue of what it knows? It would be nice to see what is already there.

No, not yet, but good idea!

Then add spore prints and we will really have something!

I uploaded a picture of my math assignment, and it showed a recommendation of mushroom!! I am not sure if one should trust this to actually figure out if a mushroom is edible or not.

Please include example images from an excluded test set for demonstration. I'd rather not have to google for pictures, download one, upload it, ...

Neat idea, however I can't even pronounce the name.

I've actually got a pretty solid name for you unregistered, but I'm not posting it publicly.

Thank you, I didn't realize it wasn't pronounceable to English speakers. Perhaps I'll think of another name (but it sounds great in French!), or you can send me your name idea at myusername@myusername.fr so that I steal it!

\ʃɑ̃.pi.ɲɔuf\

Or /ʃɑ̃.pi.ɲuf/, which I guess you could approximate in English as “shom-peen-yoof”.

Bien joué !

Here is another tool that identifies many other things http://imageidentify.com/

Well, first example was a complete failure.

https://www.imageidentify.com/result/1mvakwaegdqtn

Reply on news.ycombinator.com