Title: How Spatial Artifact Coherence Dictates Codec Resilience in Patch-Based rPPG

Abstract

While remote photoplethysmography (rPPG) demonstrates low heart-rate estimation errors on uncompressed benchmarks, its practical deployment in telehealth, neonatal intensive care, and driver fatigue monitoring relies on compressed video channels. Despite this, no previous study has identified the specific physical metric that predicts when spatial decomposition methods surpass global-projection techniques under codec compression. To address this gap, we introduce Spatial Artifact Coherence (SAC)—a metric calculated as the ratio of off-diagonal to diagonal energy within the 4x4 inter-patch Green-channel covariance matrix, filtered within the 0.75-2.5 Hz bandpass range. This metric is utilized alongside the PatchPCA algorithm family, comprising four codec-aware rPPG algorithms.

Our evaluation encompassed 280 subjects from three public datasets, testing 13 algorithms against 11 codec degradation variants, including MPEG-4, H.265, H.264, JPEG, and various chroma subsampling configurations. Statistical significance was determined using Wilcoxon tests with Benjamini-Hochberg False Discovery Rate correction (q < 0.05, 904 total tests). The results indicate that SAC accounts for 93.8% of the variance in PCA advantage between variants (r = +0.969). Distinct clustering was observed between codec families, with no overlap: non-MPEG-4 variants exhibited SAC values between 0.10 and 0.18, achieving win rates for PCA methods of 84-90%. Conversely, MPEG-4 variants clustered at SAC levels of 0.48-0.59, showing a significantly lower 61% win rate and a 5.8-fold reduction in mean improvement.

At the individual subject level, 78% of cases confirmed the expected pattern (p < 10^-22, dz = 0.73). However, the correlation of SAC at the subject level within a specific variant was weak (r = +0.099), suggesting that SAC serves to classify codec families rather than predict individual performance outcomes. The detrimental impact of MPEG-4 is structural, stemming from macroblock DCT geometry rather than noise amplitude, and is governed by the source codec state rather than video resolution. Consequently, P-Hybrid was identified as the most robust algorithm for deployment. We establish two necessary conditions for the advantage of PatchPCA: SAC must be below 0.30 and motion must be low-to-moderate, effectively ruling out raw-to-MPEG-4 transcoding pipelines. Ultimately, SAC offers a physically grounded metric for selecting codec-aware rPPG algorithms in clinical remote monitoring systems.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC