Photo crops have to keep the face and other recognition-critical regions

What it is

What people upload is often a full body, a group, a landscape frame. The avatar slot is a small square or circle. Cutting the large picture into the small slot by geometric centre or a top-third rule often keeps a shoulder, the sky, or someone else's ear. Recognition runs on the face — the small patch of eyes and features, not on the photograph as a composed picture. The crop's job is to keep the identifying region in the slot, not to fit the original "whole".

Why it happens

Face recognition leans on the eye region and the centre of the face, and barely on background, clothes, or compositional centre. An automatic crop aimed at the image centroid or a simple centre rule prefers large areas: landscape, body, several people. A circular slot is harsher than a square: the corners go, and a chin or hairline that still showed in the square goes with them. Small size then eats another layer of detail. If the slot holds only a clothing colour, the channel degrades to "the red shirt" — about as useful as a flat placeholder, while still spending the trust of "a real photo". People also upload non-self pictures (landscape, pet, logo). Then there is no face, and the rule has to fall back to another salient region rather than hunt a face that is not there and crop into empty space.

Studying it

Collect real uploads (full body, group, landscape, no-face) and contrast geometric centre, face-detection anchoring, and manual crop on "is this this person". Independent: crop strategy. Dependents: recognition accuracy, and whether both eyes sit in the slot. Stratify by upload type: a close single portrait survives most crops; full body and group shots are where the strategies split. Report detector miss rate too; the fallback after a miss is the crop the product actually ships.

Where it stops holding

• When the user drags the crop, their box wins even if the face is off-centre — some people want a profile or a back as identity.
• A no-face picture (landscape, mark) cannot "keep the face". Use a salient region or let the user pick; do not invent a detected face.
• Shape unity governs the slot's outline, not what is cut inside. When both are wrong, keep the face first, then argue circle versus square.
• In a tiny notification badge even a sharp face will not fit. Crop gains approach zero; a coarser identity mark is better.

Applying it

• Default the crop to face detection anchored on the eye region; for a group, to the person who was tagged. On detector failure, fall back to centre and offer a draggable box.
• Preview circular slots as circles. Do not pass a square crop and punch a circle later — a chin visible in the square may be gone in the circle.
• How to check: sample real uploads and see whether the shipped slot contains both eyes. The share that does not is the defect rate. Sit geometric-centre and face-anchored versions of those samples side by side; the gap speaks.

Related

• Same group: F6.09.1 Avatars identify people; defaults need distinguishable placeholders · F6.09.2 Solid-colour initial placeholders collide on same name or same colour · F6.09.3 Inconsistent avatar shapes slow list scanning
• Nearby: E4.02 List items · F3.04 Scan patterns
• Search terms: avatar crop · face-aware crop · facial recognition region