README.md

tf.keras Implementation of Hourglass Models for Human Pose Estimation

Implementation of the paper Toward fast and accurate human pose estimation via soft-gated skip connections by Bulat et al. and Stacked Hourglass Networks for Human Pose Estimation by Newell et al. using tensorflow keras.

This work has been done as part of a student course at Intelligent Process Automation and Robotics Lab (IPR).

Project setup

Initial requirements:

• Python 3.8+
• Tensorflow GPU software requirements for TF>=2.5.0

Through pip

pip install tf-hourglass-networks

From source

git clone https://github.com/benjs/hourglass_networks.git hg-nets
cd hg-nets

python3 -m venv venv/
source venv/bin/activate

pip install -r requirements.txt

Training

Start training with

python -m hourglass.training --config default_config.yaml --visible-devices 0

or in short

python -m hourglass.training -c default_config.yaml -v 0

Most parameters are adjusted in the config file and some can be changed through passing args. See python -m hourglass.training --help for more information.

Data setup

Download the mpii dataset images from their official site and the converted annotations train.h5 and valid.h5 by Newell et al. from the stacked hourglass repository. Specify the path to both parent directories in the config file.

# Paths are relative to home
mpii_annot_dir: 'mpii_annot'
mpii_image_dir: 'mpii_images'

Logging

The training process can be viewed either through tensorboard or Weights and Biases. Adjust the following line in your config file.

logger: 'wandb'  # or 'tensorboard'

Config overwriting

Add

overwrite_config: default_config.yaml

to a config file to take all parameters from another config file and update them with the parameters from the current file. The files have to be in the same directory.

Unit tests

Run all unit tests with

python -m pytest