the words, defined
Face recognition glossary.
Plain-language definitions for the face recognition and API terms you'll meet across SightRadar — from faceprint and embedding to similarity threshold and false-match rate. Every term has its own anchor, so you can link straight to it.
Concepts
- Face recognition
- The umbrella task of determining who a face belongs to. It spans verification (1:1 — is this the same person?) and identification (1:N — which enrolled person is this?). SightRadar exposes both through index, search, compare, and detect operations.
- Face detection
- Finding where faces are in an image and returning a bounding box and quality signals for each — without deciding identity. Detection is the first stage of every other operation. See DetectFaces.
- Faceprintaka face template, biometric template
- A compact numeric summary of a face derived from an image — the biometric template a recognition system actually compares. In SightRadar a faceprint is stored as an embedding (a vector), never as the original photo. A faceprint is biometric, sensitive data; treat it accordingly and only generate one with a lawful basis and consent.
- Embeddingaka face vector, feature vector
- A fixed-length list of floating-point numbers that places a face in a high-dimensional space, so that two images of the same person land close together and different people land far apart. Searching and comparing faces is really measuring distance between embeddings. The original image cannot be reconstructed from the embedding.
- Verification (1:1)
- Confirming a claimed identity by comparing one face against one reference — "is this the same person?" Powered by CompareFaces. Common in KYC and login.
- Identification (1:N)
- Finding which (if any) enrolled person a face belongs to by searching it against a whole collection — "who is this?" Powered by SearchFacesByImage.
- Enrollmentaka indexing, registration
- The one-time step of adding a person's faceprint to a collection so they can be recognized later. In SightRadar, enrollment happens through IndexFaces or selfie registration.
API objects
- Collection
- A namespace your indexed faces live in — for example, one collection per event or per tenant. Creating a collection is free; only face operations are billable. Collections are isolated: one account's collections are never cross-searchable with another's. See CreateCollection.
- Face IDaka point ID
- The identifier assigned to a single indexed faceprint inside a collection. Once you have a Face ID you can search against it directly without re-sending the image — see SearchFaces and the
search-by-idroute. - Model version
- The identifier of the recognition model that produced an embedding (e.g.
buffalo_l). Embeddings are only comparable within the same model family, so responses report the model version for traceability.
API operations
- CreateCollection
- Rekognition-compatible operation that provisions a new collection. Free to call. Maps to
POST /v1/collections. See the API reference. - IndexFaces
- Detects every quality-gated face in a photo, computes a faceprint for each, and stores it in a collection so it's searchable later. Billed per photo processed, never per face. Maps to
POST /v1/collections/{id}/index— see /docs#index-faces. - SearchFacesByImage
- Takes one image (typically a selfie), finds the largest face, and returns the indexed faces it matches, ranked by similarity. This is the headline "find every photo a person appears in" call. Maps to
POST /v1/collections/{id}/search— see /docs#search-faces-by-image. - SearchFaces
- Searches a collection using a Face ID you already indexed, instead of a fresh image — a pure vector lookup. Maps to the
search-by-idroute. See /docs#search-faces. - CompareFaces
- A 1:1 check: given two images, returns a similarity score for whether they show the same person — the building block of identity verification. Nothing is stored. Maps to
POST /v1/compare— see /docs#compare-faces. - DetectFaces
- Locates and quality-gates faces in an image and returns bounding boxes and attributes — without storing anything. Useful for upload-time validation ("is there exactly one clear face?"). Maps to
POST /v1/detect— see /docs#detect-faces.
Scoring & accuracy
- Similarity scoreaka match score, confidence
- A number (SightRadar reports 0–1; Rekognition reports 0–100) describing how alike two faces are. Higher means more likely the same person. You decide what counts as a match by setting a similarity threshold — the score itself is calibrated, not a verdict.
- Similarity thresholdaka FaceMatchThreshold
- The cut-off you choose for turning a similarity score into a yes/no match. Raise it to reduce false matches (stricter), lower it to catch more true matches (looser). In Rekognition-compatible calls this is
FaceMatchThreshold. The right value depends on your use case's tolerance for the two error types below. - False-match rateaka FMR, false accept rate (FAR)
- How often the system says two faces match when they belong to different people. Raising your threshold lowers FMR. Security and verification flows usually want a low FMR.
- False-non-match rateaka FNMR, false reject rate (FRR)
- How often the system fails to match two faces that are the same person. It trades off against the false-match rate: tightening the threshold to cut false matches tends to raise false non-matches. Tuning is about choosing the balance your product can live with.
- Liveness detectionaka presentation-attack detection (PAD)
- Checking that a face in front of the camera is a real, present person rather than a photo, screen, or mask. Liveness is a separate safeguard from recognition; recognition answers who, liveness answers is this a live capture. Use both together for anti-spoofing in verification flows.
- Quality gate
- The minimum bar a detected face must clear — size, sharpness, pose, detection confidence — before it's indexed or compared. Faces that fail are rejected rather than producing a low-quality faceprint. This is why IndexFaces reports both a detected and a rejected face count.
Billing
- Per-photo billing
- SightRadar bills per photo processed, never per face. A photo with zero faces is still a valid, charged, successful result. See full pricing.
Now put the words to work.
Spin up a key and index your first faces. Migrating from AWS Rekognition? It's a two-line change.