Distance, threshold, and status.
| Element | What it means | What it does not mean |
|---|---|---|
| MATCH | The current face-api identity-comparison readout. | It is not a combined confidence score for every model in the Lab. |
| Number | The nearest Euclidean distance between the current 128-value descriptor and a saved baseline. Lower is closer. | It is not a percentage and cannot be read as “91% escaped”. |
| / 0.58 | The current 2D match threshold. A distance at or below it is classified as a match. | It is not the detector confidence threshold, which controls whether a face box exists. |
| Status | A label derived from baseline precision, detection, the active Ghostyle, and distance. | It is not a guarantee about other systems or the next frame. |
| Sparkline | A short history of recent distances and the threshold reference. | It is not a statistical summary or calibrated uncertainty interval. |
"states" in which the lab can be.
No baseline
The camera may be working and a face may be detected, but there is no saved identity reference to compare with.
Waiting for a fresh comparison
A baseline exists, but the interface has not received a new result for the current state.
No usable live descriptor
The current pass did not return a face. This may reflect the input (dazzled/camouflaged), threshold, or any other failable step.
Distance within threshold
The nearest descriptor is close enough to a saved baseline under the current 2D rule.
Distance beyond threshold
The detector still returned a face, but the composite descriptor is so distant than a registered face that can't be considered reliable.
Composite detection absent
With a Ghostyle active, the modified-image pass did not return a descriptor. This is detection failure, not a measured identity distance.
The visible box and the match can come from different passes.
The Lab keeps a live path for tracking the original camera frame. When a Ghostyle is active, it also draws the intervention into a composite and runs another detection and comparison on that modified image.
A face can therefore remain visibly tracked while the composite pass returns a larger distance or no face. The box does not, by itself, tell you which image produced the identity result.
Detection and matching answer different questions.
| Threshold | Current default | Question | Direction |
|---|---|---|---|
| Detection score threshold | 0.5 | Does TinyFaceDetector report a face candidate? | Higher requires stronger detector confidence. |
| 2D match threshold | 0.58 | Is the descriptor close enough to a saved baseline? | Distance at or below the threshold is a match. |
| Experimental image-similarity threshold | 0.85 | Is the image embedding similar enough to the saved image embedding? | Similarity at or above the threshold is a match. |
The top Lab readout currently renders the face-api section. The Loader records both the 2D distance and experimental image-embedding similarity in its activity details. These raw values should not be combined or compared directly.
Conditions that might affect the test repetition.
- Tool and application revision.
- Input file or camera, resolution, frame rate, and controlled timestamp where available.
- Model files, browser, and inference environment.
- Baseline identity within the test, without publishing the raw biometric descriptor.
- Ghostyle, parameters, pose, distance, light, and threshold.
- Whether detection returned a face and which pass produced the displayed value.
- A series or range of observations, not only one favourable frame.