Mastering Computer Vision with TensorFlow 2.x

Understanding the hourglass model

The hourglass model achieves the state of the art results across all joints in the MPII human pose datasets, but this comes at the cost of resource-intensive network bandwidth usage. This results from the difficulty in training due to the high number of channels per layer. The FastPose distillation (FPD) was introduced in CVPR 2019 by Feng Zhang, Xiatian Zhu, and Mao Ye in their paper titled Fast Human Pose Estimation. Compared to the hourglass model, FPD results in faster and cost-effective model inference while reaching identical model performance. The key features are as follows:

• Four hourglasses (instead of eight) are sufficient to predict 95% model accuracy.
• Dropping from 256 to 128 channels results ...