Why this pool exists.
Color-science datasets live in many incompatible formats — paper-PDF tables, vendor xlsx, ad-hoc JSON, and code-only loaders that disappear when libraries change. Several widely-used datasets carry well-known but rarely-cited caveats (MacAdam 1974 is often mis-cited as the 1942 JND ellipses; COMBVD has three different reference whites; He 2022 ships chip and pair data in incompatible sheets). This pool normalizes the format, preserves the raw, and writes the caveats down where you'll actually see them.
Maintainer & related projects
Maintained by Görkem Yıldız as part of the helmlab research toolkit — a color-science library that uses these datasets for color-space training, evaluation, and perceptual benchmarking. If you want to do something with the data instead of just look at it, helmlab is the place to start.
Methodology
- Canonical CSV per dataset, under one of 26 documented schemas (see schemas).
- meta.yaml per dataset capturing: native illuminant + observer, native color space, what
dvliterally means, citation, and caveats. - raw/ per dataset — original files preserved unchanged.
- scripts/validate.py checks every canonical file against its declared schema (range checks, row counts, sentinel detection).
What this pool is not
- It is not a benchmark. It packages data; it does not define a leaderboard.
- It is not exhaustive — paywalled / out-of-print datasets (Brown 1957, Wyszecki-Fielder 1971, Huang 2015 SCD/LCD, Sanders-Wyszecki 1958, Bartleson 1960, etc.) are listed as Wanted in the README. PRs welcome.
- It is not a substitute for reading the original papers. The papers carry experimental conditions, observer briefing, and uncertainty estimates that no CSV row count can replace.
How to contribute
Open an issue or a pull request on GitHub. Useful contributions: missing datasets (especially the paywalled classics), corrections to caveats, additional metadata fields, schema extensions.