So, my tech savvy buddy wants to make an IA app that would help determine the sex of a plant either before flower or in early flower, but we need a data set to train the AI. I am looking for some collection of male and female preflower photographs. perhaps someone has this or knows where one might exist. worst case, i have to grind though the sexing threads to get a good data set. thanks dudes.
3 Likes
now that is a great idea for ai. something it is well suited for i would wager. if there is any way to tell early it shouldn’t be that hard to find.
i have been 100% on my preflower guesses in my past grow. it’s not a huge sample, but i am at 100% over the last 13 regs i ran. it seems to take about 6 weeks to sex a male.
1 Like
Hello.
No AI expert but understand the basics of image processing and reliability of results from AI.
I think this community would be a goldmine to make the project work but also caution that the accuracy of output is dependent on the input sample set (grinding).
For example, we have your pictures at 100% reliability describing M/F included in the sample set and were taken with highest resolution/ perfect focus. We also have sporadic 65% accurate M/F photos with mixed resolution. The output will be thrown off by every bad input.
Probably preaching to the choir but wanted to make the point. Need to vet all sample sets.
Hope to see progress on this.
Thanks!
John
6 Likes
i’ll be starting about a dozen or so of two or three varieties at a time and can take daily progress shots of each for you in exchange for access to the tool when you get it done.
thanks for the support.
For sure, that would have to be the deal…not sure there is an intention to monetize…so any product might just be free anyway.
Let me check with the programmer guy tonight to see what the requirements of the photos would be if any. that is a lot of work for you to do over a long time, and getting wrong would suck.
I have lots of fem pics. maybe a few males too, but having the life cycle in sets might be more meaningful.
1 Like
having it in sets will allow it to find the tells earlier. i’d make it a buck or something.
also, the son in law is a professional photographer
I dont think that this would be useful as the point where an AI would be able to determine sex would be the same point that a person could determine sex.
If there was another way besides the way we determine then it could work earlier.
1 Like
the point is to determine it before a human can with consistency. you only know the ways you know how to determine it. the AI may find other attributes or notice things we can’t. It would also shut down a few threads here an else where. And, it’s a test for a much more difficult task which I won’t reveal here until much later.
1 Like
ok, so i am not sure far off shots will work. I think you need to take photos at the crotch of the branches/nodes where the first bracts/calyx form. maybe think, what would a person submit when asking for help sexing their plant? so, close ups and maybe some a foot or two back.
The programmer was like, ‘so we need like 10,000 photos’…but i don’t think that is feasible. anyways, we just need to run two checks: confirm male or confirm female. then, there is the possibility of no confirmation, but we can cross that bridge once we get there. Thanks for your support. Getting this sorted would be pretty cool.
you can’t use anything that isn’t confirmed, period. there was a comment above that said why not. you may want to take them showing the whole plant and let ai decide what to look for, not show the ai what you think it needs. it may find a leaf structure that sexes a plant that you would never have thought of. of course if they’re too big to get the whole plant at a close enough range you’ll need to adjust, but i would think node spacing is at least as important as the crotch. if you’re getting folks to submit pictures for you that have yet to be confirmed, i would have a closeup and a shot of the whole thing just in case there is something that a closeup will miss. it’s also worth waiting until you confirm the sex so you can avoid bad data in the training set. as far as 10,000 pics, that’s a low estimate from what i know - (not much). some i’ve read about take millions of pics to train.
now all of this is from a non-expert in the field and it could all be wrong. just thoughts on my part from the limited reading i’ve done on it.
1 Like
i just read about meta’s new ai art generator that was trained on 1.1 billion pictures from intagram and facebook.
1 Like
good points. I had considered some of this. Perhaps a variety of picture styles would be good. I was saying close and maybe a bit away, but maybe there are other ideas out there. I really don’t know much about photography. maybe there is granularity i wasn’t expecting. I don’t think we need a dna confirmed dataset just yet. if we get a bad dataset, yes, the error will be higher. We can test all of this. once critical mass is reached and tests are successful, we can gain confidence. We just use another dataset to test it.
all you have to do is find out which ones were male and female. no dna anything needed. just wait a few weeks. the worst thing is getting bad data in. other than that i don’t know much about it at all.
1 Like