Quality Metrics and Measurement
Compression claims are only meaningful when accompanied by a measurement method. This page surveys the objective metrics commonly used to compare video configurations and explains where each one is and is not trustworthy.
Signal-based metrics
Peak signal-to-noise ratio compares a reconstructed frame with its source on a per-sample basis. It is cheap, deterministic and universally available, and it correlates only loosely with perceived quality. Structural similarity improves on it by modelling luminance, contrast and structure, and generally tracks the eye better than a pure error metric.
Perceptual models
Modern perceptual metrics learn their mapping from large collections of human ratings. They account for masking, temporal behaviour and the fact that artefacts in flat regions are more visible than the same magnitude of error in busy texture. The trade-off is expense and calibration: a model trained on one content population may not transfer perfectly to another.
A practical measurement loop
| Step | Action | Purpose |
|---|---|---|
| 1 | Select a representative clip set | Avoid overfitting a single scene |
| 2 | Encode at several dial values | Build a rate-quality curve |
| 3 | Score with one signal and one perceptual metric | Cross-check the ranking |
| 4 | Inspect the worst frames by eye | Catch artefacts metrics miss |
| 5 | Record conditions with every number | Make the result reproducible |
From measurement to policy
Measurement informs the ladder and the profiles described elsewhere on this desk. A score by itself changes nothing; a score attached to a configuration and a delivery target becomes an engineering decision. See the ladder planner and the 1080p profile for where these numbers land.