Understand why the same face can produce different measurements. Explore the assumptions, calculations and limits behind the result.
These five studies use synthetic coordinates or a simple camera model. They are original computational experiments, with no human participants or photo-analysis accuracy claims.
Move the camera in a simple projection model and see why nearby features look larger.
View method & results →Thirty controlled inputs, one scoring engine, and every result available to inspect.
View method & results →Hold the landmarks fixed to separate a photo-quality penalty from geometry.
View method & results →Keep face widths the same and watch the jaw angle change the closest match.
View method & results →One measurement changes, two ratios move, and an average conceals the tradeoff.
View method & results →Every table is calculated from the published inputs using the site's geometry code or the stated pinhole formula. No results are typed in by hand.
Download all inputs and results (JSON) ↓Download these four JavaScript files into one directory, then run node reproduce.mjs with Node.js 22 or newer. The printed JSON contains all five datasets and the assumptions.
Read our editorial policy for the difference between measured data, simulation and illustration. These computations test formulas; they do not validate the landmark detector on real people.
