Separate landmark estimates, metric calculations and animal selection, and understand why WhatZoo does not guarantee accuracy or personality predictions.
Separate three stages
- Landmark estimation: MediaPipe estimates facial landmarks from a photo. Google’s documentation describes an estimate of 478 landmarks. The point count is not an accuracy measure.
- Metric calculation: WhatZoo derives seven inputs from estimated x/y coordinates. Identical valid coordinates produce identical formula outputs; this does not mean the estimates have zero error against real dimensions.
- Animal selection: The app counts matches to programmed conditions. It does not identify a personality or an objectively correct animal for a person.
What the available evidence supports
The code exposes formulas and rules, and constructed inputs can illustrate candidate scores. That explains implementation, not accuracy on real people. The service does not present a ground-truth definition, independent evaluation sample and error analysis establishing animal-classification accuracy, so it does not offer an accuracy percentage.
Why the same photo can produce another result
Pose, expression and obstruction can affect landmark estimates. Even with identical measurements, a highest-score tie is resolved randomly. Identical coordinates across browsers or devices are not guaranteed, and different results cannot all be blamed on the photo.
Two candidates passing a gate and three details both score 4. That counts conditions; it does not mean 100% resemblance or personality accuracy.
How to inspect a result
Distinguish rules from fictional concepts in the encyclopedia. To explore repeatability, record the original photo, device, browser and result; liking the result is not an accuracy test. The photo guide helps inspect inputs, without promising identical outputs.
Character concepts about personality, romance or work are not assessments of real people. Read the analysis guide for the calculation sequence.