OpenPose: Real-time multi-person body, face, and hands estimation

https://github.com/CMU-Perceptual-Computing-Lab/openpose

by sytelus • 8 years ago

150 35 8 years ago

github.com

m12k• 8 years ago

This looks promising. Any thoughts on the license?

wiz21c2• 8 years ago

no commercial use it seems.

Not 100% related, but the copyright holder(s) are not clear. There is a list of authors so I'd bet they are the copyright holders. But given the fact their email addresses somewhat point to CMU, I wonder if CMU is behind...

mcsb1• 8 years ago

The Github account is from CMU perceptional computing lab. Thus I'd consider it reasonable to assume that CMU is behind.

zanalyzer• 8 years ago

The prof is also associated with oculus/fb, so the license has to be friendly to them as well.

seeekr• 8 years ago

There's actually a link somewhere to a page[0] that, all the way at the bottom states the price as USD$25,000 for a "Commercial License, yearly renewal".

[0] https://flintbox.com/public/project/47343/

SiempreViernes• 8 years ago

"You agree that all and any such derivatives and modifications will be owned by Licensor"

Assigning ownership of your own modifications to them seems significantly stronger than forbidding redistribution of your modified version. With this, just getting the examples to run on your own machine is effectively free development for the authors.

Chris20484• 8 years ago

Is that legal? You can claim ownership of derivatives??

0xfaded• 8 years ago

Paper https://arxiv.org/pdf/1611.08050.pdf

TLDR; convnet for joint localization, old school optimization techniques (graph coloring, etc) to put the pieces back together.

Re license, I don’t at all mind research being commercialized, but the use of “open” in the name rubs me the wrong way. Who in their right mind would work on this? At best it will become a benchmark to compare other systems against.

The FSF requires a snail mail letter for copyright assignment. I don’t know the specifics, but this suggests magical auto-assign copyright clauses may not hold up.

Personally I GPL my stuff that may be commercially valuable. It allows a community to maintain an open version and produce offshoots, while you maintain your own version with only code you wrote for licensing.

sandGorgon• 8 years ago

The training data comes from

Check https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimatio... for training questions. We are using their model.

https://github.com/CMU-Perceptual-Computing-Lab/openpose/iss...

What is the difference between this one and the other one ?

VikingCoder• 8 years ago

How does it do on modern dance?

https://www.youtube.com/watch?v=FOZ6KnVPvIU

lettergram• 8 years ago

How does this compare with just calling all the OpenCV functions in parallel then merging?

It kinda seems like what they did, which if it works - then kudos, but the licensing should be easy to avoid then.

svantana• 8 years ago

Errr, what? This project uses huge neural nets for detection, nothing like that is available in OpenCV. I think they only use OpenCV for IO (accessing webcams etc).

anc84• 8 years ago

What a disgrace to the "Open" prefix. Guess we need a FreeAndOpenPose instead...

grondilu• 8 years ago

Couldn't this be used to automatize fire weapon aiming?

Or worse : a laser beam aiming automatically to the eyes. It'd be quite bad since it'd be cheap and unlike fire weapons, lasers are easy to build or acquire.

How long until a mass shooting uses such technology?

shezi• 8 years ago

Realistically, if you wanted to do such a thing, you could have used a Kinect for a long time. Even face detection would work in your example, and we've had robust and cheap detection for a long time. So I assume that this is a movie-plot threat.

However, for a dramatic interpretation of your query, I suggest "Daemon" and "Freedom TM" by Daniel Suarez. The book shows groups using these technologies to great effect.

swframe22• 8 years ago

I like OpenPose. This area is progressing so quickly.

DensePose was published recently: http://densepose.org/

End-to-end Pose was published late last year: https://akanazawa.github.io/hmr/

m1sta_• 8 years ago

At a very high level how do they differ?

swframe22• 8 years ago

All of them take 2d images of humans:

open pose -> 2d points of major parts. (generation of 3d points is under development)

end-to-end pose-> 3d mesh of human body

dense pose -> 2d uv maps that can be placed on a 3d mesh

I should also mention:

AutoHair -> 3d mesh of hair from 2d image

http://gaps-zju.org/autohair/

ClothCap -> 4d cloth transfer to different 3d human model

(this one doesn't do "2d image to 3d cloth" but probably something that can be done starting with ClothCap)

Face -> 3d mesh of face

https://github.com/AaronJackson/vrn

scoot• 8 years ago

Given two views of the same scene, could the output of something like (NotActuallyOpen)Pose theoretically be combined to create a marker-less 3D motion tracking solution?

OpenPose: Real-time multi-person body, face, and hands estimation

Related Stories