CelebA Face Attributes Set

Web faces with boxes, landmarks and forty attribute tags per photo, the multi-task ground for face analysis.

CelebA pairs each face with boxes, landmarks and forty attributes, so one photo trains many face tasks

Why Does This Exist?

Face work splits into boxes, points and tags, often on different sets that cannot be compared. CelebA put web faces with boxes, five landmarks and forty binary attributes on the same photos, so face analysis could test multi-task models jointly.

This page covers multi-task contract and appropriate use. Landmark detail is in landmarks.

Think of It Like This

One school photo, many checklists

One class photo gets three checklists: circle each face, dot eyes and noses, tick forty yes-or-no traits.

Same photo, three jobs. That shared grading is CelebA.

How It Actually Works

Purpose is multi-task face analysis. Input is one web face photo, outputs are boxes, landmark points and forty attribute bits. Appropriate use is attribute and landmark baselines with bias audits. Celebrity skew limits generality, so verify on broader faces before product claims.

Toy attribute walkthrough

Take 10 photos for smiling where truth has 6 smiles. A model flags 7 smiles with 5 right. Precision is $5 / 7 \approx 0.714$ and recall is $5 / 6 \approx 0.833$ . The gap says over-flagging leads here.

Code

Attribute precision and recall on toy counts.

1def pr(hits: int, flagged: int, truth: int) -> tuple[float, float]:2    return round(hits / max(flagged, 1), 3), round(hits / max(truth, 1), 3)3
4
5print(pr(5, 7, 6))6# -> (0.714, 0.833)

Five of seven flagged with six true gives about 0.714 and 0.833.

Watch Out For

Ignoring celebrity and attribute skew

Public figures, makeup and age skew tags. The symptom is attribute scores that drop on everyday faces. Fix it by slicing by subgroup and reporting gaps, not just means.

The Quick Version

CelebA maps one face photo to boxes, landmarks and forty attribute bits.
Input is a photo, outputs are boxes, points and tags together.
Toy 5 of 7 flagged with 6 true gives about 0.714 precision and 0.833 recall.
Use it for multi-task face baselines with bias checks.
Confirm on broader faces because celebrities skew.