Quality Metrics and Measurement
A compression result is two numbers braided together: a size and a distortion. Neither means much alone. This sheet covers how the desk forms a distortion figure, how a curve is read, and how a metric can mislead.
Full-reference scoring
The desk works mainly from full-reference metrics, because a reconstruction can always be compared against the lossless source. A simple pixel-difference metric is fast and stable, and it catches gross errors such as dropped frames, shifted crops and the wrong colour matrix. It is weak on the artefacts a viewer notices, so it is never the only figure reported.
Perceptual metrics
A perceptual metric weights the error by what an eye is likely to see, so a soft gradient and a banded one score apart even at the same pixel error. These metrics are closer to preference but demand more compute and depend on viewing distance and display size. The desk states the viewing assumption with every perceptual figure, because a score without it is not portable.
| Family | Catches | Weakness |
|---|---|---|
| Pixel difference | Drops, shifts, colour faults | Blind to visible structure |
| Multi-scale structure | Blur and blockiness | Weak on temporal artefacts |
| Perceptual | What an eye prefers | Needs a viewing assumption |
Reading a distortion curve
Score a clip at three or more rate factors, then plot the score against the factor. The useful zone is the knee, where the curve stops falling quickly and starts flattening. Below the knee each step yields large visible gain for a small size increase; above it each step saves little size for a growing loss. Report the knee, not the factor.
Common traps
A metric can move without a visible change if the crop is off by a frame or the colour matrix disagrees. A metric can also stay flat while an artefact grows, if the artefact is temporal and the metric is spatial only. Always keep a short visual check beside the numbers, and never publish a score without its crop, its metric name and its viewing assumption.
