Chia-Jung Chou, Jui-Ting Chien, Hwann-Tzong Chen, Self Adversarial Training for Human Pose Estimation, arXiv:1707.02439

This code is tested with Torch7, CUDA 8.0 and Ubuntu 16.04.

Download MPII Human Pose dataset, Leeds Sports Pose Dataset and Leeds Sports Pose Extended Training Dataset. Place the images in data/mpii/images and data/LSP/images.

The file data/lsp_mpii.h5 contains the annotations of MPII, LSP training data and LSP test data.

A model trained on MPII and LSP dataset is available here.

Run the model on LSP test set and see the PCK and PCP scores. Prediction will be saved in src/evalPose/prediction

$ th main.lua -finalPredictions -nEpochs 0 -loadModel /path/to/model

Train your model with MPII and LSP training data

$ th main.lua -expID exp1

then the models will be saved in exp/LSP/exp1.

You can add the options such as -initial_Kt 0.5, -lambda_G 0.01, etc. More options can be found in src/opts.lua

Check out the fantastic repo of "Stacked Hourglass Network" for some advanced usage of this code. For example, continue previous experiment with the same or different settings.

Thanks for the open source from Alejandro Newell, this code is heavily built on his repo pose-hg-train