TPA-Flow: Physiological Tensor Interactions and Rectified-Flow Decoding for Remote Photoplethysmography

This is the authors' abstract. We don't add key points for this paper.

Remote photoplethysmography (rPPG) enables contact-free physiological monitoring with commodity cameras, providing an enabling technology for telemedicine and intelligent health systems. Existing approaches remain limited by weak physiological periodicity priors, inadequate modeling of multidimensional spatiotemporal interactions, and waveform distortion introduced by one-step deterministic decoding. We propose TPA-Flow, a unified framework that combines quality-aware physiological representation, Physiological Tensor-Product Attention (TPA), and direct-anchored Rectified Flow decoding. TPA injects a confidence-weighted periodic prior into attention and uses low-rank tensor products to capture high-order interactions across facial regions, cardiac cycles, and feature subspaces. The Direct branch first produces a stable waveform anchor. Conditioned on this anchor and a consensus of physiological candidates, the Flow branch learns a short continuous transport path that corrects residual phase and local morphological errors. The two paths are fused to preserve both periodic stability and waveform fidelity. On PURE, UBFC-rPPG, and MMPD, TPA-Flow achieves MAE/RMSE values of 0.04/0.05, 0.47/0.62, and 4.13/7.01~bpm, respectively. It also obtains the lowest RMSE in three of four zero-shot cross-dataset settings with directly comparable baselines. These results demonstrate strong accuracy, robustness, and cross-domain generalization.