2.2.0

Default match threshold changed from 0.6 to 0.5

FaceRecognitionR300Core.defaultThreshold is now 0.5, matching the reference implementation of this model.

This changes match decisions for any app that relies on the default, with no code change on your side. If you tuned anything around 0.6, set defaultThreshold explicitly on your recogniser instance before upgrading.

Alignment vertical offset changed from 0.4 to 0.35

FaceAlignmentConfiguration.verticalOffset is now 0.35, matching every other implementation of this model. The alignment regression is otherwise identical, including the -0.46 / 0.46 / 0.0 / ±0.39 coefficients and the y0 / y1 / y2 derivation, so the offset was the only divergence.

Compatibility

• Template format is unchanged. Templates are still 512-dimensional with version id 300, and templates you have already stored continue to load.
• Embeddings are not identical across this change. The same input image now produces a slightly different embedding — measured at 0.984 to 0.991 cosine similarity against the previous configuration. If you hold enrolments created with an earlier version and generate probes with this one, scores could vary by 1.2%. That is normally well inside typical decision margins, but worth considering.