Intel Movidius Neural Compute Stick Allows Students to Learn AI Programming

cageymaru

cageymaru

Fully [H]
Joined
Apr 10, 2003
Messages
22,348
The Intel Movidius Neural Compute Stick is a development kit for Deep Learning. It can be used to design, develop, and prototype smart edge devices like the image classifier shown in the video below. Other uses would be autonomous drones, smart security cameras, industrial machine vision equipment, and more. To use the device, all an aspiring AI programmer would need is the Neural Compute Stick and an existing computer. A Raspberry Pi running Raspbian OS will suffice as will a desktop or laptop running Ubuntu OS. Internet access is not needed for the device to work.

Deep Learning Made Easy Host Ian Bragg and Ashwin Vijayakumar, from the AIPG at Intel, discuss how to deploy image classifiers on edge devices with the Intel Movidius Neural Compute Stick (Intel Movidius NCS). Challenging Ian to a race against the program, Ashwin demonstrates how quickly and accurately a Raspberry Pi* with the Intelآ® Movidius NCS categorizes images. Afterward, they walk through the process for setting up an image classifier program on an edge device with a Raspbian* or Ubuntu* OS and an internet connection to download the SDK and sample code.
 
As an Amazon Associate, HardForum may earn from qualifying purchases.
Here is what a compute stick looks like....
il_570xN.1536192085_nt3l.jpg
 
bobdabilder said:
Well, what's the worst that could happen
Click to expand...
What is happening in China right now with their "social credit scoring" system and real-time facial and body recognition.
Prepare yourselves, the dark cyberpunk future is right around the corner, and we aren't going to have to wait for 2077 - 2020 will do just fine.
 
Good stuff. Anything to bring about the end of humanity and the era of cyborgs and/or superintelligent computers is good in my book. I wonder how many people share this view? There must be some, you could come to this conclusion from miserabilist philosophers like Schopenhauer, Emil Cioran, etc.
 
Cost about $112. I kind of want to get one and try it out, but not thinking of a practical use case for my house or line of work (could get the boss to buy it).
Similar tech will eventually be part of our systems.
 
Dont know how useful this would be for real time image recognition. If for example I had a usb camera and I want to recognize faces in a room, I would need a python script that
1) loops and checks for exit criterion
2) and in each loop
a) grabs a frame from a usb camera
b) downsizes the image to an appropriate filesize and draws a box around it
c) sends the image to the neural compute stick
d) gets the label from the stick
e) applies the label to the image and sends that out to a window

Even if I had a model pretrained with all the images of my family and friends, How would it be able to identify multiple images in a frame?

This example seems to run fine in real time without a NCS
https://www.pyimagesearch.com/2018/09/24/opencv-face-recognition/

but requires some good hardware.

This example runs on a pi but uses the older haar cascades
https://www.pyimagesearch.com/2015/05/25/basic-motion-detection-and-tracking-with-python-and-opencv/


Ultimately I would like to run face recognition in real time on a pi but I guess I would have to train my own model and then figure out how to make a graph to deploy it to the NCS



EDIT: I just found the NCS example
https://movidius.github.io/blog/battery-powered-dl-engine/
 
bwanaaa said:
Dont know how useful this would be for real time image recognition. If for example I had a usb camera and I want to recognize faces in a room, I would need a python script that
1) loops and checks for exit criterion
2) and in each loop
a) grabs a frame from a usb camera
b) downsizes the image to an appropriate filesize and draws a box around it
c) sends the image to the neural compute stick
d) gets the label from the stick
e) applies the label to the image and sends that out to a window

Even if I had a model pretrained with all the images of my family and friends, How would it be able to identify multiple images in a frame?

This example seems to run fine in real time without a NCS
https://www.pyimagesearch.com/2018/09/24/opencv-face-recognition/

but requires some good hardware.

This example runs on a pi but uses the older haar cascades
https://www.pyimagesearch.com/2015/05/25/basic-motion-detection-and-tracking-with-python-and-opencv/


Ultimately I would like to run face recognition in real time on a pi but I guess I would have to train my own model and then figure out how to make a graph to deploy it to the NCS



EDIT: I just found the NCS example
https://movidius.github.io/blog/battery-powered-dl-engine/
Click to expand...

Thanks for the github link, my interest in this basically matches what you want to do with the RPi and lightweight image recognition.

Also important to note for anyone interested, the intel neural compute 2 stick came out less than a month ago and does have beta support for raspberry pi now. I'd recommend looking at that model vs the older movidius (compute stick 1) as intel doesn't seem to be updating the repos specific to the original movidius stick anymore.
 
As an Amazon Associate, HardForum may earn from qualifying purchases.
You must log in or register to reply here.
Back
Top