Be it lack of physical space or our resistance to building habits, even with rising awareness, regular workouts are still difficult to achieve. Even more difficult and costly is finding someone to motivate yourself at a time of your comfort. To solve this problem, we are building a device that teaches exercise with continuous visual feedback and keeps the user engaged.
In this hack session, we will discuss different problems that need to be tackled in building such a system:
- Demo: Yoga trainer on edge devices
- Edge hardware for deep learning inference
- Model optimization:
- Quantization: Create small and fast networks that fit on the edge
- Weight pruning
- Cascade models of varying accuracy
- Real-time video feedback: A flow framework to minimize lag between the input video stream and model output. When performing exercises, it helps the user take corrective action by taking in visual feedback
- Performance measurement:
- Quality of feedback: E.g. we found in our experiments that audio feedback for exercise is not enough
- Live metrics for system performance
- Measuring devices for their compatibility
Key Takeaways:
- Challenges involved in deploying deep learning on mobile/edge devices
- Real-time feedback using deep learning models
- Use of deep learning in health and fitness domain
Check out the video below to know more about the session.
