Ask HN: How to train a neural network based on pose estimation landmarks?
I am experimenting with pose estimation. In this case, the pose estimation uses your phone camera to detect the positions of your body's joints.
Now I will film myself performing simple actions, like, say, a pushup, or perhaps throwing a punch, several dozen times. I will process these videos in order to export the frame-by-frame positions of each joint.
What's the easiest way for me to take these time-series arrays and train a small neural network on them, so that given a new video, or rather, the last N frames of a live video feed, I can detect the action that was performed, if any?
I can already imagine how to preprocess/normalize the training data. I just need someone to point me in the right direction to learn how to train a simple model and perform inference.
I am using Python.
Thanks for any help!
2 comments
[ 4.8 ms ] story [ 9.8 ms ] threadThe pose estimation converts the video of the body's movement into a time series where each frame is represented by an array of 33 landmarks (joints, basically). Each landmark is represented by normalized x, y, and z (between -1 and 1, or between 0 and 1; up to me) values, in addition to a visibility score (0-1). To keep it simple, all videos will feature a single subject taken from the same position and angle and with the subject standing in the same position and facing the same direction.
As far as classification/output, I just want booleans or confidence scores that indicate which action(s) were likely performed.
It looks like I'm looking for a tutorial on "Multivariate Time Series Classification". Is this correct?
I ended up using this tutorial: https://www.youtube.com/watch?v=PCgrgHgy26c
I manually copied every line of code before I realized someone in the YT comments uploaded their copy to github. lol.
This space and the relevant libraries are evolving so quickly that some changes were required.
All in all, this project took an afternoon. I'm shocked at how accessible this technology has become. Don't assume you can't handle it! Give it a shot if you're interested in doing something similar.
Another youtube comment mentions that it would have taken 10% as much coding if you use Keras instead. I cannot confirm or deny, but that would have been nice.