Face Embedding
A learned fixed-dimensional vector that represents identity-related facial characteristics for similarity comparison.
Representation Pipeline
A face embedding is normally not an identity vector produced directly from arbitrary raw pixels. Detection, cropping and often alignment occur first, and errors in those stages propagate into the embedding.
Comparison
Embeddings can be L2-normalized and compared with Cosine Similarity or another distance/scoring method.
Objectives such as ArcFace are designed to bring same-identity representations closer while separating different identities in angular space.
Threshold and Domain
There is no model-independent universal similarity threshold. Camera quality, resolution, pose, age, illumination and the required false-match rate influence threshold selection.
Because embeddings are biometric representations, storage, access and deletion policy also belong to system design.
Related Concepts
- Face Detection
- Face Alignment
- Vector Database