Open kevinwylder opened 4 years ago
Preliminary results are.... disheartening.
It looks like the dataset used to train PoseNet does not represent skateboarding videos. See for yourself
These results are far from being able to physically understand the wireframe of the individual, looks like we'll have to take another approach
Pose estimation is an important direction of this project, how we get to that dataset is another beast. I see two options:
I'm leaning towards option 2 because it would be nice to have all the data collection process isolated into javascript, leaving python to be purely for training our model. But doing this either expands the scope of what our model needs to learn, or pushes off the inevitability of needing to learn how to do pose estimation in python.