Dear Fellow Scholars, this is Two Minute Paperswith Károly Zsolnai-Fehér.
Pose estimation is an interesting area
of research where we typically have a few images
or video footage of humans,
and we try to automatically extract the pose a person was taking.
In short, the input is one or more photo,
and the output is typically a skeleton of the person.
So what is this good for?
A lot of things.
For instance, we can use these skeletons
to cheaply transfer the gestures of a human onto
a virtual character, fall detection for the elderly,
analyzing the motion of athletes,
and many many others.
This work showcases a neural network
that measures how the wifi radio signals bounce
around in the room and reflect off of the human body,
and from these murky waves, it
estimates where we are.
Not only that, but it also accurate enough to tell us our pose.
As you see here,
as the wifi signal also traverses in the dark, this pose estimation works really
well in poor lighting conditions.
That is a remarkable feat.
But now, hold on to your papers,
because that’s nothing compared to what you are about to
Have a look here.
We know that wifi signals go through walls.
也许 这意味着……不可能 对吧
So perhaps, this means that…that can’t betrue, right?
It tracks the pose of this human as he enters the room,
现在 当他消失的时候 看
and now, as he disappears, look,
the algorithm still knows where he is.
This means that it can also detect our posethrough walls!
What kind of wizardry is that? Now,
note that this technique doesn’t look
at the video feed we are now looking at.
It is there for us for visual reference.
It is also quite remarkable
that the signal being sent out is a thousand times weaker
than an actual wifi signal, and it also can detect multiple humans.
This is not much of a problem with color images,
because we can clearly see everyone in an image,
but the radio signals are more
difficult to read when they reflect off of multiple
bodies in the scene.
The whole technique work through using a teacher-studentnetwork structure.
The teacher is a standard pose
estimation neural network that looks at a color image
and predicts the pose of the humans therein.
就是这样 没什么新奇的 然而
So far, so good, nothing new here. However,
there is a student network that looks
at the correct decisions of the teacher, but
has the radio signal as an input instead.
As a result,
it will learn what the different radio signal distributions mean and how they
relate to human positions and poses.
As the name says,
the teacher shows the student neural network the correct results, and the
student learns how to produce them from radio signals instead of images.
If anyone said that they were working on this problem ten years ago,
they would have likely
ended up in an asylum. Today,
What a time to be alive! Also,
if you enjoyed this episode,
please consider supporting the show at Patreon.com/twominutepapers.
You can pick up really cool perks
like getting your name shown as a key supporter in the
video description and more.
Because of your support,
we are able to create all of these videos smooth and creamy, in
4k resolution and 60 frames per second andwith closed captions. And,
we are currently saving up
for a new video editing rig to make better videos for you.
We also support one-time payments throughPayPal and the usual cryptocurrencies.
More details about all of these are available
in the video description, and as always.
Thanks for watching and for your generous support,
and I’ll see you next time!