Abstract

We present BlazePose, a lightweight convolutional neural network architecturefor human pose estimation that is tailored for real-time inference on mobiledevices. During inference, the network produces 33 body keypoints for a singleperson and runs at over 30 frames per second on a Pixel 2 phone. This makes itparticularly suited to real-time use cases like fitness tracking and signlanguage recognition. Our main contributions include a novel body pose trackingsolution and a lightweight body pose estimation neural network that uses bothheatmaps and regression to keypoint coordinates.

 

Quick Read (beta)