Human pose estimation – PoseNet

TensorFlow released the PoseNet model, which is used to detect human poses using a browser. It can be used for both single poses and multiple poses.

PoseNet is based on two papers from Google. One uses the top-down approach, while the other uses the bottom-up approach.

Top-down approach

The first paper is titled Toward Accurate Multi-person Pose Estimation in the Wild and was written by George Papandreou, Tyler Zhu, Nori Kanazawa, Alexander Toshev, Jonathan Tompson, Chris Bregler, and Kevin Murphy. You can find the paper at https://arxiv.org/abs/1701.01779.

This is a two-stage, top-down approach:

• The bounding box coordinates (x,y,w,h) are determined using Faster R-CNN with the ResNet...