AI Classifier for BirdBox, train for any kind of litter. Part 2
Vložit
- čas přidán 19. 01. 2021
- Buy me a Coffe ?
www.buymeacoffee.com/Birdbox
A common question is: What will be your next step, what will be the next challenge for the birds...
To be able to classify general items like cig-butts, soda-cans, slugs, dandelions etc, i need a general purpose detector. This is part 2.
I'm trying to create this detector based on skills i have from my profession.
In this video i show you my experiments with the easy-to-use software:
www.lobe.ai
If you like my work you are welcome to support me.
www.buymeacoffee.com/Birdbox
This software brings the power of AI and machine-learning to everyone! I plan to follow up this wide with instructions on how to transfer this network to a raspberry pi
Links:
3Blue1Brown
/ @3blue1brown
This site is so god! The way he explains math has changed my way of looking at math totally.
Then...you test the networks out in real time here:
sc s.ryerson.ca/~aharley/vis/fc/
There are so many tutorials on python out there. This guy has been my teacher:
pythonprogramming.net/ - Věda a technologie
For false positive, you should normally have a confidence score. If it is less than x%, you should probably assume that the cup is empty.
Genious ! SOO looking forward new tests, it's better than a movie series !!
Well if it's not stuff like sticks or leaves being detecting, I would think it works good enough. But either way, only real testing determine if it even matters. It can be faulty with a million objects, but if the birds never pick up the million objects it's faulty in, then it works good enough.
Hey Hans,
thanks for inspiring me to build my own. I decided to try to build mine one without a micro processor.
I use 420-430 nm LEDs to light up the stuff falling down in front of a black background.
Plant parts should absorb this kind of light very well, while plastics give a normal up to a strong feedback.
That's what I use to measure with 10 light detecting sensors. If the signal is strong enough a relais will lock which has a delayed shutoff built-in.
The relais will turn a motor with an auger on for a second or so, so some food will drop down from the storage.
If this is interesting, I'll keep you updated :)
Please share this progress somewhere :)
Good Job, Hans! You're quite an engineer!
Wow, this is a really clever idea! You inspired me to create something like this myself :)
Hey mate, is there anyone i can get in touch with you? Would love to chat to you about some stuff!
I think if you and Hans work together you could make a really cool video.
Hi. I would love to chat with you, but i don't want to spread my contact details in this forum...can i reach you in some way ?
/Hans
@@LabbHasse Have a look (on a notebook): czcams.com/users/Ididathingabout. He got a mail address in his youtube profile :)
NO WAY HES DOING THIS
Hope it will works fine, and i am looking forward to watching birds using it ! Very good project 👍🏽
My friend, you are a genius! 😄
This is amazing man, well done
I just discovered your channel, and I love the bird feeder. I also love the idea to use AI to detect objects... but I'm guessing you don't know about nVidia's Jetson? it's like a Raspberry Pi, but with an nVidia GPU capable of processing AI without sending data to external servers, and it's capabilities make it perfect for a project like yours. The basic model, the nano, It's not terribly expensive, costing about 65€.
I hope you find this information useful!
Well, the classification do happen on the Rasberry Pi :)
No external servers involved.
@@RubenKelevra Yeah, but the GPU on the Raspberry Pi is both underpowered and not designed to compute AI tasks. The Jetson product line is specifically designed to do AI. This means easier, faster and more accurate results.
@@Spyd77 but it comes with a cost: much higher consumption. Dunno if it's worth here. You just need an motion sensor and a couple of frames analysed. Speed works great in this demonstration.
@@RubenKelevra Jetson Nano and Raspberry Pi 4 are similar on power draw. But the key here is efficiency. Using both systems for AI, will result on the Jetson winning on power economy just because it's made for that, where the Raspberry Pi has to crunch a lot harder to get the same (or mostly worse) results.
@@Spyd77 well, running a dedicated GPU all the time has definitely drawbacks on power consumption. The 2-3 minutes per day when the system has to analyze images since there was a motion detected doesn't really matter much :)
thank you very much for this video, and thanks to the youtube algorithm for recommending the 3 bottlecaps video in the first place. this lobe ai might be exactly what i was looking for in a project.
edit: yep, this worked perfectly for my project.
The details of the build are far too technical for me, but I admire you for using your intellect for such a serendipitous endeavor 💕
Hi are you still doing these experiments
You are a genious
Increible
Let just bring the birds real-world stuff and train on this, to optimize the precision. The bird's training will just follow up. I doubt there's any "damage" to the training when they bring some stones one day and get rewards and this stops working some days later :D
Hi Hans, this is amazing, where i can get one?
Hello ! That's so clever ! What system will you use to "eject" the thing they will bring without scaring them so they can get more butts & cap ?
I’m planning some mechanical solution, maybe powerd by a RC-servo tja Will ’open up’ the bottom of the Bowl an sort litter into different containers...i’ve Done similar designs before. I’ll keep you updated....
@@LabbHasse I would like to help if possible. I'm used to 3D print, using servo motor, and Fusion 360. Would you like to let me work with you ?
Would it be possibel to position the camera below the platform and let it capture the deposition area from below? You could basically make an mirror set up in 45° to see everything from above and the trapp door for the litter conncected at an motorized hinge. So when the camera detects something right it dropes down into an garbage collection and if its detects something wrong it tillts up to throw the leaves or stones off. There will be the probelem of lighting if the camera looks from below but maybe it will simplyfy the algerithm training as it will go more for shape than color. Or you could use an camera with can capture in a different light spektrum as procesed matierals and natural stuff behave really different under ir or uv light. And to better train the birds you could basically create an spektrum of rewards if he camera detekts litter but there is some debri in it so it dispenses a lower reward and for pure drops or multipel drops it gives higher rewards. If you want to go kompletly insane you could set up an second camera which trys to reconizes each bird and reward them more for conesquitiv or constant litter deliveries and lower the reward amound after so many times per day or even randomize them to hool the birds more
I was thinking pneumatics
How much litter have the birds actually picked up?
If you take this too far, birds in you area will start stealing cigarettes out of people's hands and mouths :) you should experiment with "snusbokser" (container for a type of tobacco normal in Sweden and Norway).
Suverän ide, här i Örebro har vi massor med kajor. :)
Kajor, jackdaws might be even smatter.
Hi. Is the code an open-source?
It would be interesting for me just to read :)
Thanks
and the 3D Printed parts (and maybe even a Bill of Materials (BOM) )
I think if everyone with access to a Raspberry Pi, a 3D Printer, and some parts was able to make this there could be a serious impact on litter ! Also all those people could aid in the development / come up with related ideas !
i want mine to collect money
ahahahahhaah
I need a video for those of us who don't know how to use these programs/pi/lobe/ECT. Shalom you're loved 💔