search

benja-wu / CozmoGestureRegonize

using 3DCNN model by Keras to recognize four type gestures, and then perform related actions
31 stars 7 forks source link

About prediction #1

Open w530248323 opened 6 years ago

w530248323 commented 6 years ago

I'm doing the same job, I also trained jester_dataset used c3d. I saw you use robot to predict the gesture, do you have any idea on how to predict data in real time on the RGB camera?

benja-wu commented 6 years ago

@w530248323
Actually, I am working on this topic too. Recently I come out with two ideas. In this toy project, cozmo costs over 100ms and under one seconds for calculating the result . What if we cut this period down to 10ms? Firstly, I would like to rewrite the prediction function with caffe which is build by C++. Secondly if your hardware is powerful enough( with high GPU calculation ability, which i don't have in my macbook) I recommend predict with tiny amount captured pics. Like 10 pics predict once. Filling the missing required pics ( the amount differ to the original picture number that in your traing process, in my project it is 40.) with the first frame or last frame. Hope for helping your problem.

w530248323 commented 6 years ago

Thanks a lot, my Wechat id is w530248323. We can study this topic together~

benja-wu commented 6 years ago

@w530248323 You're welcome! I wrote a article in Zhihu about these toy projects. (url is here) . Maybe you can find out some interesting details in it. And then we can discuss more about the project through Zhihu after following each other. :-)