[MUSIC PLAYING] Hello. I’m Martin Crawford. And this is the
IoT Developers Show where we look at IoT
technology, share learning opportunities, and showcase
cool demos and the creators behind them. In this episode, we talk
with Stuart Christie, an IoT evangelist here at
Intel, as he walks us through two demos showcasing the
Movidius Neural Compute Stick. Stuart, thank you so much
for joining us today. Do you want to tell us a
little bit more about yourself? Sure. I’m an IoT evangelist,
like yourself, Martin. And I’m the hardware guy
in this software group. On a personal side, I’m
interested in robotics and photography. And this is sort of
a photography example here we’re going to be showing. Fantastic. Well, I’m really
excited in this demo. Do you want to tell us
a bit more about it? Sure. It’s a follow on from
last month’s example where we were doing the
unboxing of the Upsquared Board. This time, we are
showing some demos that you can run on this board
using these Movidius Compute Sticks. These are also development
devices from Intel, much smaller, as you can see here. They’re USB devices. There’s an SDK
available if you want to write your own programs. But today, we’re
just going to demo some of the prepackaged
applications that you can find on GitHub and
the Movidius Application Zoo. Sure. So kind of like monkeys? Yes, zoo as in animals in a zoo. That’s correct. In fact, they call
it a park-like area in which deep learning models
are housed for exhibition. I’m not sure why
they call it a zoo. But a quick search for
GitHub on zoo in Python shows a lot of deep
learning examples. Excellent. So what kind of applications
are housed in the zoo then? We don’t have any monkeys. But we do have some of birds. But today, we’re going
to demo two of them. We’ll feature you
and me in the zoo using a USB camera to
do object recognition. And then the second example
shows the scalability of these USB sticks. And we’ll run the same
demo side by side, one running on one stick
and one on the second, so you can see the
speed difference. And so all of these demos are
running on pretrained models? Exactly. The first one using the
webcam has been trained on a selection of objects. It’s fun to get it confused. There are Windsor
ties, lab coats, and other seemingly random
objects that show up. This demo doesn’t do
any background removal, processes the whole video frame. So we’ll likely see
some odd guesses. One of my favorites
I’ve been trying to get is to recognize my face. And then I cover it up so
it looks like a ski mask. It thinks it’s a ski mask. Well, let’s see what it
recognizes here right now. I’ve got a bulletproof vest on. It sees me as a
person, that’s good. Oh, I’m there as a person. Interesting. So I’m assuming
that you can also tweak the threshold
levels to kind of get it to be more exacting. Exactly. This one’s actually
running a threaded example. So it’s running TinyYOLO
on one of these USB sticks. And then the second one
is really GoogLeNet. And you can tweak the thresholds
for it for both of those things using the keyword. OK. And YOLO is you only look once. And that is being used
for object detection. And then the GoogLeNet is
being used as an object classification essentially. Exactly. It refines the models, yes. There are 20 things that
the YOLONet can detect. OK. Well, do you want to take
a look at the next demo? Yes. We’ll close this down. And then we’ll start
it up and see it with these two sticks
versus the one stick. So this is the
parallel compute demo. And you’re saying
that it’s using three of these Movidius sticks? Exactly, two sticks
here versus one stick. And it shows how the
system is scalable. This demo process is a system
of directory of images. There’s about 65 images. And it scans an image, uses
Open CB to print the object name that it’s finding of the image. And then it also says running
one stick, running two sticks. And you can see this one, the
frame rate’s a lot faster. The scrolling here, the text,
says stick one and stick two there. Again, one stick, two. And then this one’s
running as stick zero. So this shows the scalability. You can add as many
sticks as you really need for your application. OK. And can I add any
photos to this database, or is it only trained
to recognize these 65? And No. You can add as many
photographs as you want. I’ve run it on my own
directory of photographs. So that works fine. But obviously, they’re trained
to recognize certain devices. So if I’ve got a
cathedral photograph and they haven’t been
trained to find a cathedral, it’s not going to find it. It may say that it’s a
porcupine or something. Porcupine or a dog, yes, because
it does do probabilities. This is what it thinks it is. So how is it actually doing
the parallel processing? It’s simply doing a ping pong,
alternating between the sticks, as all three sticks are
loaded with the same detection algorithms. I’ve seen other demos where
the controlling program pulls the sticks and looks
for a free stick to use. That’s a better algorithm to
use, for example, if the time to recognize something
is indeterminate, from, say example,
a live video feed. OK. Is there a limit to how many
sticks you can use in parallel? Not really. A USB-3 port can support for. I believe that’s the
bandwidth limit on the port. We’ve got to photograph we
can show you on the screen right now. it’s got eight. It’s a bit of a problem to move. It has an external PCI to USB
card running on an extender. So it’s unwieldy,
but you can use it. OK. So you mentioned that this
is a development platform. Now, are there any kind
of production versions of this thing out there? There are. It’s shipping in the Google
Clip camera and the DJI drone. So there’s the chip itself. There’s also a
hardware developer kit, which is a bit more like
a regular PCB for doing some software development. I think this is a great platform
for developers that want to get into computer vision. And thank you so much for
coming down here and showing it off to us. Well, I’ve enjoyed
showing it off. Thanks very much
for inviting me. In February, Intel
Technologies will help fans experience the
Olympic Winter Games, Pyongyang, 2018 with the latest
innovations in virtual reality, 5G, drones, and gaming. Check out the links
to learn about Intel at the Olympic games. Thanks for watching the
IoT Developers Show. Don’t forget to
like and subscribe. And remember to check out the
links provided to learn more. Thanks, guys. [INTEL THEME SOUND]


7 Comments

Rick Ooi · March 23, 2018 at 3:34 am

Hi, is there any limitation in image size use for training, like 25Megapixel? and is it posible for Movidius to process large image size. thanks.

Agata Fors · September 3, 2018 at 5:10 am

Для чего нужен этот предмет??? Нужно больше данных!!!

oh my gosh ponies · November 15, 2018 at 4:23 am

Can you do an update for the 2nd generation stick

Andres · November 27, 2018 at 5:25 am

2:26 what da little mistake hahaha: a person detected like a bulletproof!! (that robot will shoot you)

TransistorSnack · February 2, 2019 at 7:16 am

0:50 "where we were doing the unboxing of the of [sic] squidward"

Ludwig Schreier · May 2, 2019 at 9:05 am

"tweak" to get it "more exact" :-)))))))))))))))))

not sure · May 3, 2019 at 12:23 am

starts at 0:45

Leave a Reply

Your email address will not be published. Required fields are marked *