This is the code for my 2016 COMP3710 project on human pose estimation. To run it, fire up Matlab and use one of the following:
>>> demo_mpii % Train and test on MPII Cooking Activities
>>> demo_flic_piw % Train on FLIC and and test on Poses in the Wild
>>> demo_h36m_upper % Train and test on Human3.6M, using only upper body
Datasets are not included with this archive. However, they will be downloaded
automatically by the demo_*
scripts.
Some important caveats to note:
cnn_model.json
(network specification) and cnn_model.h5
(weights) files
for the fully convolutional network in the cache. You will likely want to use
the IPython notebook at keras/upgrading-convnets.ipynb
to do this.Since this code is being submitted for assessment, I need to document where it came from. In a nutshell, all code in this directory was written by me, this semester, with the following exceptions:
cy/
directory is adapted from Chen & Yuille's Articulated Pose
Estimation by a Graphical Model with Image Dependent Pairwise Relations
paper.cmas/
directory is adapted from Cherian et al.'s Mixing
Body-Part Sequences for Human Pose Estimation.get_deps
and the flow
sub-directory, everything in ext/
is a
third-party dependency.