This work proposes a light network for user activity guidance (such as yoga) in real time on computationally low-end CPUs.

• It is built on top of Lightweight OpenPose by Daniil Osokin.
  

The trained network is used to build a system that guides a user to learn a certain yoga by breaking it into states and providing corrective instructions throughout the activity.

The trained lighter network is capable of retaining enough information to correctly identify different yoga poses while reducing the inference time by up to 4 folds.

This is a collaborative work by Aashish Adhikari, Dhruv Jawalkar, and Sudha Ravi Javvadi.