Convolutional Pose Machines

Shih-En Wei, Varun Ramakrishna, Takeo Kanade, Yaser Sheikh, "Convolutional pose machines", CVPR 2016.

Contact: shihenw@cmu.edu.

Before Everything

• Watch some videos.
• Install Caffe. If you are interested in training this model on your own machines, consider using our version with a data layer performing online augmentation. Make sure you have done make matcaffe and make pycaffe.
• Copy caffePath.cfg.example to caffePath.cfg and set your own path in it.

Testing

• Run get_model.sh to retreive trained models from our web server.
• 1. CPM_demo.m: Put the testing image into sample_image then run it! You can select models (we provided 4) or other parameters in config.m. If you just want to try our best-scoring model, leave them default.
• 2. CPM_benchmark.m: Run the model on test benchmark and see the scores. Prediction files will be saved in testing/predicts.

Training

• Run get_data.sh to get datasets including FLIC Dataset, LEEDS Sport Dataset and its extended training set, and MPII Dataset.
• Run genJSON(<dataset_name>) to generate a json file in training/json/ folder. Dataset name can be MPI, LEEDS, or FLIC. The json files contain raw informations needed for training from each individual dataset.
• Run python genLMDB.py to generate LMDBs for CPM data layer in our caffe. Change the main function to select dataset, and note that you can generate a LMDB with multiple datasets.
• Run python genProto.py to get prototxt for caffe. Read further explanation for layer parameters.
• Train with generated prototxts and collect caffemodels.