Face Recognition with Micropython (K210 chip boards: Sipeed Maix series, M5StickV, HuskyLens)
HTML-код
- Опубликовано: 12 дек 2019
- Face recognition on a microcontroller, faster than ESP32-WHO, for almost the same price.
A review and look at the code for Micropython face recognition on K210 boards.
The board for this video was kindly provided by Seeed studio, check out this board and other hardware for machine learning acceleration at Seeed studio store!
www.seeedstudio.com/Sipeed-MA...
Implementations of Mobilefacenet(I don't know why I said "Facemobilenet" xD):
github.com/JaydenGG/Mobilefac... Tensorflow
github.com/JaydenGG/Mobilefac... Keras
Download the models at:
www.maixhub.com/
Add me on LinkedIn if you have any questions:
/ dmitry-maslov-ai
Credits for music:
• Dreams - A Synthwave Mix
Dreams - A Synthwave Mix (Juno Dreams - Paradise) Наука
Thanks for the video!!!
But, is it possible to do face recognition using esp32 camera and esp32 (as it would be cheaper than k210)?
Yes, try this one github.com/espressif/esp-who
I'd say it's not the question of price, but availability really... K210 is out of stock currently
Thanks for the video. do you ever try convert the yolov5 pt model to kmodel? I use nccase toolbox,but don’t make it.
No, I haven't. The most recent yolo that was ported to K210 is yolo v3.
Love t-shirt. You should discuss goldfinger module as cofusion exist
Wow, I had no idea it's called "goldfinger module" - you are talking about M1N SoM, right?
@@Hardwareai yes. Many are confused. That aside on their discussion board many do not know how to use the m1n, never making the connection it is the same thing sans lcd.
Do you know approximately the price to use sispeed (MaixPy/components/kendryte_sdk/libs/ ) in a commercial app? The website is all in Chinese! BTW, considering the support/documentation might not be in English, how much better is this K210 risc-v chip versus running a similar face-recognition application on other hardware.. like, JeVois (linux quad a7) or a pi/beaglebone? Or is the advantage in the software library they've created?
I think the downside might be that you can't easily test using other machine machine-learning software.. tensorflow vs pytorch?
The libraries are free to use from what I understand. They charge for face recognition models, about 1 USD for copy, I can forward you to their sales manager if you'd like.
For most of(90 percent I would say) documentation they have English version... I think main advantage is size and price - if you use chip or module, it's cheaper than pi/beaglebone and has smaller footprint.
Tensorflow works good with nncase converter, Pytorch support was added very recently and is in alpha stage.
@@Hardwareai can you please send me the sales representative contact
thanks for this video, it's really helpfull ,
do you have any idea where i can find dataset for landmark training ?
You can use this article as a starting point
towardsdatascience.com/face-landmark-detection-with-cnns-tensorflow-cf4d191d2f0
Good luck!
@@Hardwareai thanks , looks good, another question , do you have any idea of how i can perform a liveness (anti-spoofing) mechanism on the k210 or any other cost efficient edge computing ?
I would say the same approach as with other things - DNN. Of course more are available, check out this Github repo, second section(right after the picture):
github.com/imironica/liveness
There are a number of approaches to liveness detection, including:
Texture analysis, including computing Local Binary Patterns (LBPs) over face regions and using an SVM to classify the faces as real or spoofed.
Frequency analysis, such as examining the Fourier domain of the face.
...
How can I train the landmark model in a compact size currently I've following several tutorials but the size of the model is bigger than K210 memory even using pruning techniques I can't fulfill the memory requirements
Pruning cannot reduce the size of the model that much, it is more like trimming the fat measure. You just need to take an existing feature extractor, that is known to work on K210 (for example MobileNet v1) and then add the landmark detection layer on top of it.
Do you think the K210 could run pose estimation?
Yes, it can. The easiest way in my opinion is to do it using two stage inference - on first stage run person detection, then cut the person image, resize it and feed to pose estimation network.
This was awesome, but their models go a little beyond being closed source. Their keygen also permanently disables jtag? Any thoughts on that? I'm kinda pissed.
Well, as I mentioned they're a pretty small company and want to make some money from software. If it's just for personal use, it is free however. So do you need it for personal project or commercial applications?
@@Hardwareai Personal. My only point was that using their licensed things removes some functionality of your hardware, permanently. Made me think twice.
Nice video. please give your email . need to discuss some image processing problems.
I see you added me on LinkedIn already:)
@@Hardwareai Yes! however, email is a better option
don't use python but C++
don't be so mean to python(micropython) :) it's good for prototyping and learning
@@Hardwareai micropython fails every time) It is very object-or. lang so it is basically useless for deeper learning.
I tried - didn't like it. Real learning starts from C