13 comments

[ 3.9 ms ] story [ 39.7 ms ] thread
This is really awesome,this will stimulate the movie industry.
The movie industry already has these tools. This'll really stimulate indie games, if they can now do AAA mocap.
Fantastic! Thanks for another amazing Deep Learning paper to study and play with!
how can you play with this without any data?

"The training data used for the deep neural network was generated using Remedy Entertainment’s in-house capture pipeline based on a cutting edge commercial DI4D PRO system [Dimensional Imaging 2016] that employs nine video cameras."

You can always take tested ideas and apply to your own rig. The pioneering work was done by NVidia already, you see the results and it's up to you how much inspiration can you take from their work.
There are lots of sources for mocap data if you know people in the film, games, or entertainment industry.
We see a similar realtime application in snapchat[1] -- probably not as detailed as this one but might be a similar tech,

1. https://techcrunch.com/2015/09/15/snapchat-looksery/

I don't think there's any mention of NN to create those filters - in fact, not sure if it's ML at all. Very likely some off-the-shelf CV face-detection - especially considering it must run real-time on a phone.
Even the "simplest" face detection algorithms rely on strong machine learning concepts, eg: Viola-Jones uses AdaBoost to train the face classifier cascade.
Amazing if they can manage to get this level of quality without requiring very special lighting conditions.