Skip to main content
Quality Indexing measures a corpus of video and publishes the measurements over an API. It does not publish a verdict on the footage. What counts as good depends on what you intend to train, so you get the counts and the denominators and decide for yourself.

​
The model

Three nouns carry the whole API.

Dataset

A corpus of footage, addressed by id. Every id you pass to the API comes from GET /api/quality/datasets, which lists the ones your credential can read.

Clip

One span of one source video: the unit everything is measured over, and the join key between the endpoints. A clip’s identity survives renumbering.

Measurement

One thing a dataset can answer for. A dataset carries a measurement once the work behind it is stored; the listing says which of the four it has.

​
What the API answers

The first three are corpus-level reductions. clips is the same underlying measurements unreduced, so you can re-total them over any subset you care about: one operator, one site, one week of collection. GET /api/access sits beside all of them and answers what your credential can do. It never errors on a missing or bad one, which makes it the right first call.

​
What it does not publish

The endpoints publish results and the populations they were computed over. They do not publish how the results were produced — no thresholds, no model or tooling identity, no agreement rules, no scoring formula. That boundary is deliberate, and it is not withholding for its own sake. A reader who could vary the cutoffs could re-partition a corpus into whatever answer suited them, and a number nobody can re-derive to taste is worth more than one anybody can. Where a rule matters for comparing two answers, it is published as an opaque version token instead: same token, directly comparable; different tokens, not.

​
Conventions

Counts and seconds, not shares, wherever a denominator is yours to choose. Corpus-level endpoints do publish shares, because at corpus level the denominator is unambiguous — and each one arrives beside the population it was computed over. An absent field is not a zero. A clip with no frame_quality was never decoded; a clip whose frame_quality.active is 0 was decoded and found to contain no usable frames. Averaging the first as though it were the second quietly drags your numbers down. See Quality metrics. Every metric response echoes the run that answered it. You address a dataset, never a run, and you always get its most recent measurements. The echoed token is how two answers a week apart can be told apart: the same token means the corpus was not re-measured between them. Handled errors are JSON, carrying a machine-readable code and a human message. Branch on code; the message is for people.
A malformed query string is rejected before the handler runs, so it returns a 400 with a plain-text body (Failed to deserialize query string: missing field \dataset“) rather than the JSON shape above. Parse defensively: check the status first, and do not assume every non-2xx body is JSON.
This reference is generated from the server’s OpenAPI 3.1 document, which the server produces from its own routes and is verified against on every build. If the reference and the API ever disagree, that is a bug — the document cannot drift.