Books as physical objects
Every book is an object before it is a text: a coloured rectangle of a particular height and thickness, standing next to others. No catalogue records that, so Spine manufactures it — 889,495 editions so far, each measurement carrying where it came from and how far to trust it.
The tops are ragged because the heights are real. No cover art is drawn here — every stripe is a measured colour and two measured lengths.
- 01Take the spine edge, not the cover
The leftmost sliver of a front-cover scan is the part of the object that faces out on a shelf. Averaging the whole cover gives you the colour of the artwork, which is a different question.
- 02Convert to CIELAB
sRGB with a D65 white point and the 2° observer. Distances in RGB do not match distances in the eye, and every later decision is a distance.
- 03Cluster, deterministically
k-means with k = 5 and seeded k-means++, single-threaded so the arithmetic — and so the last bits of every centroid — is the same on every run. Clusters under 5% of the pixels are noise, not colours.
- 04Name the winner
The nearest bucket under CIEDE2000, with the achromatic gate widening as the spine darkens: a photograph of a black book is full of measurable chroma that nobody can see.
- 05Write it down with its provenance
Source, method version, confidence and evidence. Nothing is overwritten; a stronger source simply wins the next resolve, and the weaker one stays on the record.
- Verified
- ≥ 0.95. A person held the book and measured it.
- Estimated
- 0.50–0.95. Manufactured from good evidence — a cover's spine edge, a uniform imprint, printed dimensions.
- Guess
- Below 0.50. A weak inference, often wrong, and labelled so.
- Unknown
- Nothing is known. Not a guess — a gap.
Verifying a spine means holding a colour card against a physical book; verifying a thickness means a ruler across it. Nobody has done that here. The ceiling is 0.90, and it belongs to imprints whose every spine is the same — Penguin Classics is black because that is what the imprint is, not because a camera saw it. That empty tier is the most truthful thing on the site.
Open Library's image for a book is whatever someone uploaded. For some editions it is a scan of the title page, so the measured colour is the colour of paper — correctly measured, from the wrong thing.
An ISBN names an edition; the image attached to it may be a reissue in a different jacket. The measurement is right about the picture and wrong about the book.
Most editions have no published thickness, so it comes from the page count: pages ÷ pages-per-inch × 25.4 + a binding allowance. That is 20–30% out by construction, and it runs high on long books, because publishers use thinner paper exactly when a book gets fat.
Searching by bucket takes under 100 ms. Searching by an arbitrary colour evaluates CIEDE2000 against every candidate in a provable lightness window, and over 686,000 editions that takes 400–630 ms at the median depending on load, against a 300 ms target. The one acceptance gate currently red.
What Spine is
Spine is an application of VER, a standard for verifiable records about images. Every cover it measures has a record: its content hash, its canonical pixel hash, its perceptual hash — enough to prove which image a colour came from, without Spine ever keeping or republishing the image. Cover art belongs to its publishers; here it is hotlinked from Open Library, and only hashes and measurements are stored.
A record that validates means it is well-formed, never that it is true. A claim that verifies means its producer said it, never that the book is that colour. The interface is built so that saying otherwise would take deliberate effort — which is the point.