BlazePose: On-device lightweight convolutional neural network architecture for human pose estimation

Every physical activity we do today can be tracked using Fitbit, phone apps, etc. Every step taken can be counted, but are these values accurate. Moving our hands can be considered as a step at times by our phone apps. For all the fitness enthusiasts, Google has introduced BlazePose for tracking human posture. The approach provides human pose tracking using machine learning to infer 33, 2D landmarks of a body from a single frame.

What sets BlazePose apart from other pose models
1. Accurate localization of more key points, thus making it more suitable for fitness applications.
2. Achieves real-time performance on mobile phone with CPU inference

The pose estimation is done two-step detector tracker ML pipeline. Using detector pose region-of-interest(ROI) is first detected. The tracker then predicts 33 pose critical points from the ROI. For video mode, the detector is run only in the first frame.
BlazePose has a massive application in fitness and yoga trackers using body posture.

Github: https://google.github.io/mediapipe/solutions/pose

Paper: https://arxiv.org/abs/2006.10204