Face recognition enrollment from still images: what it takes

A face template built from a still image is only as strong as the frame it came from. A guided live enrollment checks pose, distance and lighting while the person stands at the camera and simply takes another frame when one fails; an imported photo gets no second chance, so every quality gate has to be checked by hand. The numbers below define those gates: the 80-pixel minimum between the pupils, the ±15° pose window, the compression and blur budgets, and a worked pixel budget for a 600 × 600 identity-photo crop. With them you can decide, file by file and batch by batch, whether to import, re-shoot or enroll live.

What a still-image enrollment actually stores

A face template — the data some vendors call a faceprint — is a compact vector of measurements extracted from face imagery; matching compares vectors, not pictures. Guided live enrollment runs a feedback loop: the device finds the face, checks its size in pixels, its pose and the light on it, and keeps capturing until the frames qualify. The stored template then inherits that checked quality.

Image enrollment removes the loop. One frame goes in, the algorithm extracts whatever it can, and the template stores exactly that quality. The decision rule is simple: if a frame would have passed every live check on its first attempt, it can enroll from a file; anything weaker produces a template that generates today and generates false rejects tomorrow.

The five gates a still must pass

Each gate below is one a live flow enforces automatically; on a still, you measure them yourself.

• Pupil distance. 80 pixels between the eyes is the minimum for reliable template quality and liveness support, and 100 pixels or more is recommended for smoother matching. Treat anything under 80 as a reject.
• Pose. Near-frontal, within about ±15° in yaw and pitch, eyes level. Past that angle the far side of the face foreshortens and the template skews toward the visible cheek.
• Sharpness. No motion or focus blur. At 100% zoom the eye corners and lashes stay crisp; if they smear, the small features the template needs are gone.
• Exposure and light. Not overexposed, not too dark, evenly lit across the face, with no hard shadow across one eye and no specular glare on forehead or nose.
• File and compression. jpg, png or bmp; keep face-region compression between 20:1 and 24:1, the compromise a NIST face-image format study recommends (Understanding the face image format standards). Do not re-save or upscale files; both destroy real detail.

Worked example: the pixel budget of a 600 × 600 identity-photo crop

Identity documents standardize the portrait, which makes them a good test case for the pixel gate. Inputs:

• Crop size: 600 × 600 px minimum, 1200 × 1200 px maximum, printed at 2 in × 2 in — the composition standard identity documents use.
• Head height on that print: 25-35 mm, set by the same standard.
• Adult proportions: crown-to-chin about 230 mm against a pupil gap of about 63 mm, so a head is about 3.6 pupil gaps tall (an estimate from average adult anthropometry).
• The gates: 80 px minimum, 100 px recommended pupil gap.

The rule in words: convert the head height to pixels at the crop's scale, then divide by 3.6.

Arithmetic:

• Scale: 600 px ÷ 50.8 mm = 11.8 px/mm (the same pixels-to-mm conversion used in coin size sorting with a single camera).
• Smallest allowed head: 25 mm × 11.8 px/mm = 295 px. Largest: 35 mm × 11.8 px/mm = 413 px.
• Pupil gap: 295 ÷ 3.6 ≈ 82 px; 413 ÷ 3.6 ≈ 115 px.

Result: a minimum-size identity crop lands between 82 and 115 px of pupil gap — over the 80 px floor everywhere, but at the recommended 100 px only when the head fills most of the frame. The threshold works out to a head at least 360 px tall: 360 ÷ 3.6 = 100 px, and 360 px is 60% of the 600 px frame. The same crop at the 1200 × 1200 maximum doubles every number: 590-826 px of head, 164-229 px of pupil gap, comfortable margin. Upsampling a 600 px file to 1200 px doubles the count but adds no detail, so it buys nothing.

Decision: accept a 600 × 600 file only when the head fills at least 60% of the frame height; otherwise ask for the 1200 × 1200 original or re-shoot.

What a still cannot prove: liveness

A guided live enrollment has a witness: the depth-and-motion sensor watches a live face at the right distance while the frames are taken, so the template carries proof of presence. An imported file has none — a scanned document photo is as still as a phone selfie. That is why the 80 px pupil minimum is stated for reliable liveness as much as for template quality: anti-spoofing needs the same facial detail the matcher does.

If your platform supports image enrollment, read what its API reference says about anti-spoofing for file-enrolled users before you rely on it. For doors and machines where a wrong accept costs money or safety, keep live enrollment and use imports only to pre-provision records, then re-enroll in person. Depth hardware built around capturing 2D and 3D images together from one scene is what carries that liveness channel, and the 3D depth sensors category is where those parts live.

The documented path when the SDK has no image import

Some device lines ship image enrollment as an API entry point; others announce it for a future SDK release. Match the feature to the version: an entry point absent from the API reference you have may exist in a newer release, so check the release notes and the vendor's own SDK distribution first. If no version exposes it, the supported route is the vendor's support channel and roadmap — not unpacking the library, replaying internal calls or reconstructing the feature from binaries.

The same discipline applies to the templates themselves. A face template encodes one algorithm's internal features; no cross-format with open-source descriptor libraries exists, and descriptors from one library cannot be matched by another. When you change algorithms or device lines, plan a re-enrollment of the population, and during any overlap expect the two systems' match scores to differ.

Sizing a capture station for enrollment stills

When you capture the stills yourself — a badge desk, a kiosk, a station on the line — the gates become machine settings: fixed working distance with manual focus, a global shutter for a subject who sways, a color sensor, and even frontal light you control.

The pixel requirement is lighter than most people expect. A 100 px pupil gap at a 63 mm adult pupil distance needs 100 ÷ 63 ≈ 1.6 px/mm on the face, and a head filling 60% of a 720p frame is 432 px tall, or about 120 px of pupil gap — the pixel gate clears on any sensor from 720p up. Resolution is not the bottleneck; focus, pose and light are. Set the frame first, then confirm the pixel count:

Three hardware patterns cover it: a compact color sensor with manual focus for a fixed desk mount, a higher-resolution Ethernet sensor for head-and-shoulders crops, and a variant with an integrated light for even frontal illumination.

The MV-SC6050C-08M-XBN combines 5 MP color resolution, a global shutter, an 8 mm M12-mount lens and an integrated multi-color light source in one housing. The MV-SCB007EC-12S-WBN covers the desk-mount pattern: 1.6 MP color, a 12 mm manual-focus M12 lens and Fast Ethernet output. For lighting, hold exposure and gain fixed once the station is lit: auto exposure and auto gain on color cameras shows why letting the camera re-balance per frame destabilizes a gated pipeline, and lighting for face recognition covers the drift problem on the matching side. Both the vision sensor finder and the camera configurator filter by shutter, resolution and mount.

When a still is the right input

Import works when the photos already meet the gates: a corporate portrait program shot at one station with controlled light, recent badge photos from the same camera, or high-resolution originals you can crop to the 60% head rule. It also fits pre-provisioning — loading records before people reach a camera — as long as live re-enrollment follows.

Import fails on document scans (small faces, off-angle, laminate glare), video-call snapshots and group photos, where the face region usually sits below the 80 px floor. Frame grabs from a video pipeline are their own middle case: exposure processing differs from a true still capture, which video exposure and frame rate: why still images look brighter explains, so validate a grab against the gates rather than assuming it behaves like a photo.

Before you import a batch

Work the gates as a per-file checklist and log the result for each person:

• Measure the pupil gap; reject under 80 px, prefer 100 px or more.
• Confirm the head fills at least 60% of the frame height.
• Check pose within about ±15° of frontal, eyes level.
• Zoom to 100% on the eyes; reject any smear.
• Check the face for blown highlights and deep shadow.
• Confirm jpg, png or bmp, no upscaled files, face-region compression in the 20:1-24:1 band.

Each gate maps to a failure you would otherwise meet months later as a false reject with no image to debug. Route every failed file to live capture instead of forcing the import, and keep presence-checking live enrollment at any door where presentation attacks are realistic.

The decision, per file: measure the gates, import what passes, re-shoot or capture live what does not. For capture hardware, start in the smart cameras category, and the 3D vision topic page collects the capture-side guides in depth. If the station needs a lens, light or mounting decision alongside the camera, the configurator walks the rest of the system; for a template or liveness question specific to your device line, the contact page reaches the applications team.