420chan now has a web-based IRC client available, right here
Leave these fields empty (spam trap):
Name
You can leave this blank to post anonymously, or you can create a Tripcode by using the float Name#Password
Comment
[*]Italic Text[/*]
[**]Bold Text[/**]
[~]Taimapedia Article[/~]
[%]Spoiler Text[/%]
>Highlight/Quote Text
[pre]Preformatted & Monospace text[/pre]
1. Numbered lists become ordered lists
* Bulleted lists become unordered lists
File

Sandwich


Community Updates

420chan now supports HTTPS! If you find any issues, you may report them in this thread
Pose estimation and rigging by Priscilla Singerstedging - Sun, 08 Apr 2018 05:39:32 EST ID:xDpPYEfL No.37493 Ignore Report Quick Reply
File: 1523180372573.jpg -(42267B / 41.28KB, 889x500) Thumbnail displayed, click image for full size. 42267
facebook released a new method to detect multiple bodies in a video. Some poeple already made their own attempts on github to do the same, and the algorithm will detect human movements and infer a skeleton out of it (you may have seen it, th eone with different colored limbs). What is the best pipeline to use this algorithm to extract movements from people in videos and place their sksletons into say anime figures in a video rendering software so the figure can follow the pose extracted form the human?

just asking for which software / language is best for this. i already know the theory, just need the practicalities of it (mostly model rigging and video stuff)

trying to make a hatsune miku x donald trump dance video but too lazy to rig it myself so i want to use AI for that.
>>
Polly Worthingforth - Sun, 08 Apr 2018 16:33:19 EST ID:Xm/W+3lL No.37495 Ignore Report Quick Reply
You'd need a way to convert from the automatically-inferred skeleton to the target skeleton. I'm not sure which software would be best for this, but most of the skeletal rigging code I've seen was in a C-like language (C++, Java, C#, etc.) and used in videogames.
>>
Phoebe Mucklespear - Tue, 10 Apr 2018 11:31:22 EST ID:9QSfnS0r No.37507 Ignore Report Quick Reply
Your best bet is python imho

  • You have tensorflow, the status quo on deep learning neural network.
Most probably any existing pose estimation networks you'd find on github will be done in tensorflow.
  • You can use python scripting in blender.
Although it uses it's own interpreter installation and it's not straightforward to use it in an outside application the things you can do in regards to animation and scripting in blender are second to none.
  • Makehuman
An application written in python to create humanoid meshes. It has the usual sliders to change the model attributes. You might try to get your network to come up with the sliders for a similar mesh based on an input video.

The advantage of this approach: You can at any point edit anything by hand and go beyond scripting.
The disadvantage: Many components and you pretty much have to at least adapt an existing tensorflow model.
>>
Phoebe Mucklespear - Tue, 10 Apr 2018 11:39:39 EST ID:9QSfnS0r No.37508 Ignore Report Quick Reply
That said:

>too lazy to rig it myself so i want to use AI for that.
At current time there is no ready to use end user application that can do this.
And writing such application yourself is certainly more work that animating by hand.
>>
A_Wizard !cMZsY.BCnU!!vVWR8L52 - Wed, 11 Apr 2018 20:43:18 EST ID:mJDH+xt3 No.37514 Ignore Report Quick Reply
>>37493
This is actually a bit disturbing. An amazing toy, but the ability to log this kind of biometric data can be used for some pretty damned nefarious ends.
>>
Samuel Berringtit - Fri, 13 Apr 2018 12:02:59 EST ID:9QSfnS0r No.37516 Ignore Report Quick Reply
>>37514
I disagree, at least in an ideological sense.
I think it's disturbing that the law still regards video as hard, tamper-proof evidence. This should have been fixed in the early 00s not now when it's becoming a trivial exercise in terms of labor.
And if you find that scary see what can be done in conjunction with statistics.
For instance they demonstrated an algorithm that can guess a persons sexual orientation with 90%+ accuracy....
Just based on the intricacies of a facial expression on one photograph.


Report Post
Reason
Note
Please be descriptive with report notes,
this helps staff resolve issues quicker.