Recovery-Aware Learning for Remote Photoplethysmography During Post-Exercise Heart Rate Recovery

Remote photoplethysmography (rPPG) provides a contactless way to estimate cardiac activity from facial videos, but post-exercise recovery remains difficult because the heart rate changes rapidly and motion artifacts are often pronounced. This work focuses on rPPG-based heart rate estimation under these non-stationary conditions. We introduce a Recovery-Aware Learning framework in which a subject-specific recovery time is used only during training to guide window sampling and auxiliary supervision. Early recovery windows are sampled more frequently to reduce the imbalance between the short rapid-transition phase and the longer stable phase, while a soft recovery-progress target provides additional temporal guidance. We further propose ResPulseNet, a 1D convolutional encoder–decoder network designed to reconstruct rPPG waveforms from multi-ROI RGB traces while sharing temporal representations with the auxiliary recovery-progress task. On the MCD-rPPG dataset, ResPulseNet achieves an overall MAE of 2.20±0.36 bpm and an MAE of 2.48±0.37 bpm on post-exercise recordings, compared with 4.89 and 5.41 bpm, respectively, for the strongest evaluated baseline. These results show that combining recovery-aware training with dedicated temporal modeling improves rPPG estimation during post-exercise physiological transitions.