Live Repetition Counting

Requirements

Python 2.7.7+

Python modules

• theano 0.8.1
• h5py
• tornado
• nose
• opencv-python

Troubleshooting

• When installing theano, tornado, nose if you come across the error that a package is already installed use the --ignore-installed flag e.g. pip install --ignore-installed theano.
• To install the specific version of theano, please run pip install theano==0.8.1.

To run live repetition counting

Weights can be found at this link. Save the latest epoch (200) as weights.save in the live_count directory.
To open the camera and do exercises in real-time, run python live_rep.py.
To use an existing video file, run python live_rep.py -i <videofile>.

To train the model

• Train model by running python2.7 rep_train_main.py in the trainNet directory. Please create empty directory weights in root directory of the repository before starting.

Data

Please find the data at this link. This is the simulated data generated by the Matlab scripts in the syn_data directory. Please have the out directory in the root directory of the repository. You only need the h5 directory for training the data, but we have included the mat unprocessed data for reference.

Major changes from original implementation

• Input data is in files with extension .h5 not .gzip.h5.
• Input to CNN is of type theano.tensor.tensor4 not theano.tensor.matrix.
• Input and output directories of data, specifically train_dir, valid_dir, weights_dir in rep_train_main.py.

Original Implementation Citation

Theano implementation by the authors of the original work can be found here.

@inproceedings{levy2015live,
  title={Live Repetition Counting},
  author={Levy, Ofir and Wolf, Lior},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={3020--3028},
  year={2015}
}