Human pose estimation via Convolutional Part Heatmap Regression

This repository implements a demo of the Human pose estimation via Convolutional Part Heatmap Regression paper Bulat&Tzimiropoulos.

Note: New code capable of running on devices with limited resources, for Human Pose Estimation and Face Alignment was released. For a demo please check: https://github.com/1adrianb/binary-human-pose-estimation

Requirement

• Install the latest Torch version
• Install python 2.7 using the package manager

Torch packages

• nn
• cunn or cudnn (preffered) if you have a CUDA enabled GPU
• optnet
• xlua
• image
• sh
• fb.python

Most of the listed package can be installed by simple running

luarocks install [packagename]

For sh and fb.python packages please visit their github repositories and carrefully follow the instruction provided by their authors.

Python packages

• numpy
• matplotlib - required for plotting

Trained models

By default, on the first run the scripts will attempt to automatically download the models, however for your convinience they are provided also for separate usage.

Dataset used	LSP error	MPII error
MPII	-	89.7
MPII + LSP	90.7	-

Usage

The provided code comes along with a series of options. In order to list them please run:

th main.lua --help

To run a demo on 10 random images:

th main.lua -dataset lsp 

To evaluate the model on the validation set for LSP/MPII:

th main.lua -dataset lsp -eval

If you have installed cudnn4 or cudnn5 you can run the demo faster:

th main.lua -dataset lsp -eval -usecudnn

The demo doesn't require a GPU, however having one will speed up the process.

Notes

For more details/questions please visit the project page or send an email at [email protected]

Warning: The script will download by default both the models and the dataset(~15Gb), if you wan't to avoid this or you already have them downloaded please move them in the corresponding folders in datasets/[datasetname]_dataset/. Running the demo for lsp dataset will require only ~700Mb of space.