See Face_Parsing_2016/matlab/face_parsing/run_demo.m for details. Please download the data_db from MultiHelen_v1.mat.
Update:
For original dataset, please find the download link here: HelenDataset.
Per request, I am adding a similar python version HERE. The python version should be compatible with many newer caffe versions (one caffe_dev is included). All parameters are the same except a slightly different face cropping strategy. Note that the python version uses IOU instead of f-score in the test.py, you need to modify it to compare with SOTA methods.