392234 Project: Deep learning Methods for Heart Rate Estimation (Pj) (WiSe 2024/2025)

Description: Remote photoplethysmography(rPPG) methods have recently gained prominence for their ability to extract heart rate from videos of the face. However, these methods often lack robustness in real-life scenarios due to motion and varying illumination. Recent work by Paruchuri et al. [1] has demonstrated that employing motion transfer as a data augmentation technique improves state-of-the-art results by 79%. This project aims to validate these implementations and apply the model to a dataset collected using the Digital Stress Test (DST) paradigm [2], which induces stress in participants.
Tasks:
• Validate the results obtained by Paruchuri et al. [1].
• Train a deep learning model and evaluate its performance on the DST dataset. Required skills:
• Basics of neural networks and machine learning.
• Proficiency in Python.
• Familiarity with PyTorch. Supervisor(s):
• Bhargav Acharya, <bacharya@techfak.uni-bielefeld.de> (primary contact)
• Prof. Dr. Hanna Drimalla, <drimalla@techfak.uni-bielefeld.de> References:
[1] Paruchuri, A., Liu, X., Pan, Y., Patel, S., McDuff, D., & Sengupta, S. (2024). Motion Matters: Neural Motion Transfer for Better Camera Physiological Measurement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 5933-5942).
[2] Norden, M., Hofmann, A. G., Meier, M., Balzer, F., Wolf, O. T., Böttinger, E., & Drimalla,
H. (2022). Inducing and Recording Acute Stress Responses on a Large Scale With the Digital Stress Test (DST): Development and Evaluation Study. Journal of medical Internet research, 24(7), e32280.