Public methodology · Algorithm version 1.0
How WhatZoo Works: From 478 Face Landmarks to an Animal Match
WhatZoo does not send a face to a generative AI to guess an impression. MediaPipe Face Landmarker produces normalized coordinates in the browser. We convert selected points into distances and ratios, then apply deterministic profile rules. This page mirrors the current production calculation flow.
Last technical review: July 13, 2026
The seven measurements we actually use
The classifier does not treat all 478 points as independent features. It combines representative eye, nose, mouth, and face-outline points. Ratios reduce the effect of image resolution, but pose and expression can still affect results.
| Measurement | Landmarks and calculation | Meaning |
|---|---|---|
| Eye slant | Y differences of 33·133 and 362·263, summed × 100 | Positive rises toward the outer eye; negative falls |
| Eye-distance ratio | Distance 33–362 ÷ face width 234–454 | Inner-eye spacing relative to face width |
| Face aspect ratio | Height 10–152 ÷ width 234–454 | Larger values indicate a relatively longer face |
| Nose-width ratio | Distance 102–331 ÷ face width | Nose width normalized to face size |
| Mouth-width ratio | Distance 61–291 ÷ face width | Mouth width normalized to face size |
| Chin ratio | Distance 17–152 ÷ face height | Relative length below the mouth |
| Smile flag | Compare mouth center 0·17 with corners 61·291 | Simple check for raised mouth corners |
The three-stage classification flow
1. Critical-trait gate
Each animal profile first checks a required criticalTrait. For example, a tiger requires eye slant above its threshold to enter the strict candidate set.
2. Detail score
Each candidate has three detail rules. The gate contributes one point and matched details add up to three more, for a maximum strict score of four.
3. Soft fallback and ties
If no strict candidate exists, profiles are compared by detail matches without the gate. When several candidates share the top score, one is selected, so borderline measurements can produce different matches.
Worked example with fictional measurements
These values do not represent a real person. They show how normalized ratios are derived from coordinates.
- Face height 0.60 ÷ width 0.42
- Aspect ratio 1.43
- Nose width 0.060 ÷ face width 0.42
- Nose ratio 0.143
- Chin length 0.105 ÷ face height 0.60
- Chin ratio 0.175
Where are photos and coordinates processed?
Landmark detection and metric calculation run in the user’s browser. WhatZoo does not upload or retain the analysis photo or its 478 coordinates on its server for classification. User-initiated share images are also rendered in the browser.
Important limitations
- Non-frontal poses, occluded features, and hard shadows can distort coordinates.
- Expression matters because the smile flag is one of the rules.
- This is not a clinical, biometric, or psychological instrument and has no certified accuracy claim.
- Personality, romance, and career descriptions are entertainment interpretations—not scientific diagnoses from facial appearance.
Versioning and change policy
This publication describes algorithm 1.0. When metric definitions or profile thresholds change, we will update the review date and explain the material change here rather than silently tuning rules to alter result distribution.