Choosing tools

How to compare image-search tools without inventing an accuracy score

A useful comparison records known matches, misses and false matches for a clearly described test set. A handful of successful examples does not justify a universal accuracy percentage.

Construct a fair small test

Use your own images or an explicitly permissioned collection. Include an unchanged published image, a crop, a lower-quality version and an image with no expected public match. Write down expected source URLs before running the searches.

Record more than the best result

  1. Use equivalent inputs and access levels where possible.
  2. Note date, account tier and any restricted index.
  3. Count useful source links separately from similar thumbnails.
  4. Preserve misses and misleading results in the worksheet.

If one tool finds a known portfolio page and another finds a repost, both may contribute useful information. Calling only one “correct” depends on whether your task is finding the creator, finding all reuse or locating any copy.

Publish the limits with the result

Describe what your sample represents and what it does not. Separate observed timings from a guarantee about future requests. Do not imply you conducted a comparison when you only reviewed documentation. A transparent task-based recommendation is more useful than a precise-looking score without a reproducible method.

Continue with tool selection worksheet.

Published by Facehunch with AI assistance. How we use sources and maintain these guides.
KEEP EXPLORING

Related guides

FROM READING TO EXPLORING

See how a photo search comes together.

Try the fictional example, or run a consented API test. Paid searches are disabled.

Open Facehunch