Mastering Computer Vision with TensorFlow 2.x

Hourglass training

The hourglass network is trained on a FLIC human pose dataset with 5,000 images (4,000 for training and 1,000 for testing) and a MPII human pose dataset with 40,000 images (28,000 for training and 12,000 for testing).

Note that in this book, we are not using MPII datasets to train the hourglass model. The information that's provided about the MPII dataset is included to explain how the hourglass model was trained for human pose estimation.

The average accuracy of all the joints reaches about 70% in about 20,000 training iterations, with a max accuracy of around 80%.

So far, we have discussed OpenPose and the stacked hourglass methods of pose estimation. In the next section, we will discuss PoseNet.