Remote Heart Rate Estimation from Facial Videos with Balanced Contrastive Learning
Oral presentation
Research approach
Extracted global and local pulse representations with a 3D CNN. Balanced contrastive pairs align global and local representations within a video and separate global representations from different videos.
Results
Intra-dataset MAE/RMSE: 0.13/0.36 bpm on PURE and 1.08/2.49 bpm on UBFC-rPPG. The study used 75% of subjects for training and 25% for testing, excluding videos with heart rates of 135 bpm or higher.
All publications