HiDURPA (High Definition Uncompressed Remote Photoplethysmography Acquisition) is a high-definition raw video dataset designed for research in remote photoplethysmography (rPPG). The dataset includes recordings from 11 participants (4 female, 7 male), with five 30-second sessions per participant.

Participants were positioned approximately 1 meter in front of the camera and were instructed to stand still during acquisition. Video was captured using a Blackmagic Pocket Cinema Camera 4K and illuminated with two softboxes to ensure consistent lighting. The recordings were acquired via a Blackmagic DeckLink Mini Recorder 4K capture card in uncompressed YUV422 at 30 fps.

To provide reliable reference signals for validation, the following ground truth data were recorded: ECG using a Polar H10; BVP and SpO2 waveforms using a VTM20F.

The dataset will be released upon publication of the paper “A Study of Resolution-Aware Video Compression for Remote Photoplethysmography”. Access is restricted to research groups and research institutions and requires a signed Data Usage Agreement (DUA).

This section provides supplementary material to complement the paper, presenting the complete tabulated results and accompanying figures from our experiments on resolution- and compression-aware remote photoplethysmography (rPPG). The tables report rPPG performance (mean ± standard deviation) across different video codecs (H.264/AVC, H.265/HEVC, H.266/VVC, and AV1), quantization parameter (QP) settings, and spatial resolutions on the HiDURPA and UBFC1 datasets. The figures summarize rPPG performance (median ± interquartile range) across codecs, QP levels, spatial resolutions, and datasets.

 

The HiDURPA master recordings are captured as uncompressed YUV 4:2:2 (10-bit) at 1080p, 30 fps. During encoding, the four codecs in the FFmpeg-based compression pipeline use the following pixel formats:

 

Among the five rPPG methods we evaluate (GREEN, ICA, CHROM, LGI, and OMIT), the ≈4 % of sequences for which no valid heart-rate estimate could be extracted are specific to ICA. ICA (JADE) begins with a whitening step that eigendecomposes the per-window 3×3 RGB covariance and rescales the signal by the inverse square roots of its eigenvalues. Under strong compression (aggressive quantization, 4:2:0 chroma subsampling, inter-frame skip/merge coding, and downscaling), the minute pulsatile chrominance variation is suppressed, so the three colour channels become dominated by a common luminance component and nearly collinear. The covariance then becomes effectively rank-deficient (its smallest eigenvalue approaches zero) and the whitening rescaling becomes numerically ill-defined, which causes the eigenvalue-inversion step to fail. The other four methods require no such inversion (GREEN selects the green channel directly, CHROM applies a fixed linear colour projection, LGI uses the leading left singular vector, and OMIT a QR decomposition) and produced no numerical failures on the same sequences. We emphasize, however, that this refers only to the absence of numerical breakdown: under such heavy compression the estimation accuracy of all methods degrades, so the greater numerical resilience of the projection-based methods does not by itself imply accurate heart-rate recovery.

 

The dataset comprises 11 adult participants (7 male, 4 female), each recorded in five 30 s clips (55 clips total) at 1920×1080, 30 fps, under stable, diffuse illumination, with synchronized ECG as ground truth. Among the seven male participants, five were aged 20–30, one 30–40, and one 40–50. Among the four female participants, three were aged 20–30 and one 30–40. Overall, 8 of the 11 participants were aged 20–30, 2 were aged 30–40, and 1 was aged 40–50.