Video quality glossary: PSNR, SSIM, VMAF, BD-rate

No terms match your filters. Try a different keyword or clear filters.

FFmpeg

FFmpeg is the open-source media engine that decodes, scales, encodes, and pipes frames between filters. It hosts libvmaf (and the psnr and ssim filters) and runs most scripted, automated quality measurement - the default free way to measure video quality.

Fidelity

Fidelity is how exactly the delivered video matches the source signal it came from. It is not the same as how good the video looks, since invented detail can please the eye while lowering fidelity.

Freeze / freezing

Playback halting on a frozen frame because the player's buffer ran dry (rebuffering) or a decode failed. Every frame on screen is correct; the fault is in time, not pixels, so picture metrics cannot see it. A top streaming-side artifact.

Full-reference metric

A metric that compares the impaired video against the pristine original. PSNR, SSIM, and VMAF are full-reference and need the source to compute.

Gibbs phenomenon

The signal-processing law behind ringing: rebuild a sharp jump from a limited band of frequencies and it overshoots by a fixed proportion (about 9%) however many you keep. Adding frequencies makes the ripples tighter, never shorter.

Golden reference

A fixed, known-good clip and its expected metric scores, used to detect quality drift over time in a regression test suite.

Ground truth

Ground truth is the human verdict on video quality, the subjective Mean Opinion Score, against which objective metrics are trained and judged. When a metric and a properly run subjective test disagree, the ground truth wins.

Harmonic-mean pooling

A pooling method that weights low (bad) frame scores more heavily than the arithmetic mean, so a few poor seconds pull the summary score down.

Hidden reference

Slipping the unimpaired source clip into the test unlabelled, so the viewer rates it without knowing it is the reference. Used in ACR-HR to catch viewer bias and compute a DMOS by subtracting each viewer's reference score from their processed-clip score.

ITU-R BT.500

The ITU-R recommendation defining subjective assessment of television picture quality, including viewing conditions and methodologies like DSIS and DSCQS.

ITU-T P.1203

ITU-T P.1203 is the first standardized model of end-to-end adaptive-streaming QoE — the quality of a whole session, not one clip. It fuses video, audio, and an integration module folding in loading delay, rebuffering stalls, and bitrate switching.

ITU-T P.1204

An ITU-T model series for predicting the quality of adaptive HTTP streaming, with bitstream, pixel, and hybrid variants.

ITU-T P.808

The ITU-T recommendation defining a framework for crowdsourced subjective quality testing, with reliability controls like gold standards and attention checks.

ITU-T P.910

The ITU-T recommendation defining subjective video-quality test methods (ACR, DCR, PC) for multimedia applications.

Judder

Jerky, uneven motion from frame-rate conversion (such as 3:2 pulldown of 24fps film to 60Hz). A per-frame metric cannot see it.

Just-noticeable difference (JND)

The smallest change in quality a viewer can perceive. Useful as a unit — a metric difference below one JND is invisible to people.

libvmaf

The open-source library that computes VMAF (and PSNR/SSIM), available as an FFmpeg filter, making perceptual quality measurement scriptable and free.

Luma

Luma is the brightness channel of a video, the Y in YCbCr. The human eye is far more sensitive to brightness detail than to color, so quality metrics and codecs weight luma heavily and preserve it at full resolution.

Macroblock

The square pixel region a codec processes as a unit - 16x16 in H.264. It is the cell of the blocking grid, the term behind the synonym macroblocking. Modern codecs use larger variable units (HEVC CTUs, AV1 superblocks), so the generic term is block.

Mean Opinion Score (MOS)

The average quality rating from a panel of human viewers, usually on a 1-to-5 scale. The most common ground-truth measure of perceived quality.

Mean squared error (MSE)

Mean squared error: the average of the squared per-pixel differences between an original and a compressed frame. It is the raw error measure PSNR is built on — PSNR is simply MSE rescaled against the peak pixel value and put in decibels.

Mezzanine

The high-quality master source a pipeline encodes from, with quality headroom to spare. It is the true full-reference original every encode is compared against, so its quality sets the ceiling a metric like VMAF can ever measure.

Mosquito noise

Shimmering, moving speckles around moving edges across frames — the temporal cousin of ringing, named for its insect-like flicker.

Motion feature

VMAF's temporal input: how much the picture moves between neighbouring frames — in the open-source package, the mean co-located pixel difference on luma. Since the eye forgives more distortion in fast motion, it tunes how harshly spatial damage is judged.

MS-SSIM

Multi-Scale SSIM — an extension of SSIM that evaluates structure across several resolutions, usually predicting perceived quality better than single-scale SSIM.

MSU VQMT

MSU VQMT (Video Quality Measurement Tool, Moscow State University) is a dedicated multi-metric application computing around two dozen metrics including VMAF, with a per-frame plot, frame inspection, a difference heatmap, GPU acceleration, and HDR variants.

NIQE

NIQE (IEEE Signal Processing Letters 2013) is a classic no-reference, opinion-unaware image-quality model: it scores how far an image deviates from natural-scene statistics fit to a corpus of pristine photos, needing no human-labeled training data at all.

No-reference metric

A metric that judges quality from the impaired video alone, with no original to compare. The only option for live and user-generated content.

Objective quality

Quality as computed by an algorithm (a metric) — a fast, repeatable proxy for human opinion that can be wrong on content it was not validated for.

Outlier ratio

The outlier ratio is a validation statistic: the fraction of clips where a metric's prediction error is larger than the human uncertainty - specifically outside the 95 percent confidence interval of the MOS.

Pair comparison (PC)

A subjective-test method where viewers see two versions and pick the better one. Very sensitive, but the number of comparisons grows fast with conditions.

PCHIP interpolation

PCHIP is a monotone piecewise-cubic interpolation that cannot overshoot its data points, used to fit rate-quality curves for BD-rate. It replaces a plain cubic spline, which can wiggle above or below the real data and invent quality that was never measured.

Pearson correlation (PCC)

Pearson Linear Correlation Coefficient — measures how linearly an objective metric's scores track human MOS. Higher is a better-aligned metric.

Percentile pooling

Percentile pooling summarizes a clip by a low percentile of the per-frame scores - the 5th-percentile is the value below which the worst 5 percent of frames fall - so the worst moments are not averaged away. It is the number a quality gate should watch instead of the mean.

Perceptual quality

Perceptual quality is how good a video looks to the human visual system, regardless of any original. It can diverge from raw signal fidelity, as when an AI upscaler invents plausible detail that pleases the eye but was never in the source.

Per-frame plot

A per-frame plot is a metric-over-time line plot, score on the y-axis against frame index on the x-axis, exposing short dips a single pooled number hides. Plot the harmonic mean and 5th percentile beside the mean, and beware a truncated y-axis.

Per-shot encoding

Optimizing the encode recipe per shot or scene rather than per whole title, choosing the best resolution and quality setting for each shot. Finer than per-title; saves about 30% bitrate at equal quality but multiplies encodes per title 100-fold or more.

Per-title encoding

Choosing a custom bitrate ladder for each asset based on its content and a quality target, instead of one fixed ladder for everything.

Playback buffer

The playback buffer is the holding tank of decoded video a streaming player keeps ahead of the playhead, a few seconds to tens of seconds. Its depth trades startup speed against rebuffering: a deeper cushion resists stalls but slows the start.