A minimal repository (including pre-trained models) to demo the pain detection model proposed in the paper titled Unobtrusive Pain Monitoring in Older Adults with Dementia using Pairwise and Contrastive Training. It is available under the "Early Access" area on IEEE Xplore, and it can be cited with its DOI.
The code is tested on Linux with Python 3.6+, PyTorch 1.6, and Cuda 10.2
After installing the requirements, you should be able to run test.py, and it should print out the pain score (PSPI) for the frames in the example_frames folder.
Two pretrained models are included. One was trained on the UNBC-McMaster Shoulder Pain Expression Archive dataset and the University of Regina's Pain in Severe Dementia dataset. And another checkpoint that was trained on the UNBC-McMaster dataset only. In both cases, UNBC subjects 66, 80, 97, 108, and 121 were excluded from training.
test.py can also do a “frame rate test” and print out how many frames per second your system is capable of processing.
We achieved ~9FPS on a PC with an NVIDIA RTX-2080 Ti GPU and Intel i9-9900K CPU @ 3.60GHz.
Currently Face Alignment Network (and S3FD) are used to detect and align faces.
These could be swapped for faster non-deep learning methods to improve performance.