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
Some important caveats to note:
- CNN training and evaluation requires a GPU with CUDA support and a significant amount of memory (12GB K80 and K40 GPUs were used for experiments).
- Training is also taxing on disk space and memory. Several hundred gigabytes of disk storage are required for the CNN training set, and some stages of the pipeline can take gigabytes of main memory.
- The training and testing process typically takes several days, most of which
is spent training the CNN and structural SVM. You can save some time by
interrupting CNN training part way through, but you will have to manually
upgrade the partially trained model to a fully convolutional network and put
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.ipynbto do this.
- Due to its high resource requirements, I have only been able to find one machine on which to test this code. Thus, it may break if moved to another machine.
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:
- Code in the
cy/directory is adapted from Chen & Yuille's Articulated Pose Estimation by a Graphical Model with Image Dependent Pairwise Relations paper.
- Code in the
cmas/directory is adapted from Cherian et al.'s Mixing Body-Part Sequences for Human Pose Estimation.
- Except for
flowsub-directory, everything in
ext/is a third-party dependency.