Colour-Channel Selection for Video-Based Respiratory Rate Estimation: A Diagnostic Study with a Single-Shot Held-Out Evaluation

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

Remote respiratory-rate (RR) estimation from video is usually framed as a modelling problem and reported as one number; we treat it as a measurement problem. Using COHFACE (39 subjects, 160 clips) and VitalVideos (39 clips), both seated, we compare untrained colour channels over facial and torso regions with trained two-stream models on subject-disjoint five-fold splits, with held-out rows scored once. An untrained torso chrominance channel reaches a mean absolute error (MAE) of 1.267 breaths/min (r=+0.795), while every facial channel stays at noise level (3.69–3.91) against a 2.644 same-row blind baseline. Background patches average 4.36 and a neighbouring region 2.58, so the signal is not illumination but region-specific. Motion from sub-pixel displacement fails alone (2.92), and removing it from the colour signal does not degrade the estimate (1.156): the mechanism is narrowed, not excluded. Per-window selection lowers fusion error from 3.439 to 2.463 but stays worse than carrying a person’s last measurement forward (1.438, on 70% of rows) and no better than the blind baseline; with overlapping windows, that test is descriptive only. The comparison covers the projections, regions and estimator tested here.