Systematic apparent-age screening suggests the under-18 filtering (paper §3.1) has measurable false-negative leakage

[kaanyor]

Hi maintainers,

We're building a camera app with a strict internal policy against processing images of minors, and EasyPortrait is one of our aggregate-statistics sources (photometric/framing bands only — no image redistribution, no identity embeddings computed or stored). Before relying on the dataset we wanted to independently verify the paper's stated filtering (§3.1: "checking for children under 18, naked people, religious signs, and watermarks").

What we did: apparent-age visual screening (Claude-based multi-perspective review — an adversarial "try to refute under-18" pass plus an independent anatomical/facial-maturity pass) over the full locally-available train+val pool (34,000 images / 7,866 unique user_ids from meta.tsv). This is a full census of that pool, not a sample.

Headline finding: roughly 9% of individuals were independently assessed (2+ reviewers, high confidence) as more likely under 18 than over, plus a further ~18% we couldn't confidently place at 18+ either way. A handful of representative examples (age range = independent reviewer estimates, not a claim of certainty):

image_name	reviewer-estimated age range	confidence
ce08b4ce-88b0-4e7d-9af9-5dc1ad803706	early-mid adolescent	high (this is the case that originally triggered our review)
91dbdf31-32cd-4c8e-b212-b29b0f3411f6	~13-17	high
ebdb8c2b-6b99-4184-9e31-ba4c34b9b16f	~6-9	high
58490a0b-3dc1-447a-88f6-aa41bcee1cdf	~10-14	high

Separate data-integrity finding: in about 20 accounts, the frames attributed to one user_id clearly depicted different individuals — in a few cases a young child alongside an adult in the same account's photo set (e.g. f976f194-24a7-4886-b4b4-e2ab4662d7ae, one of several frames under a single user_id where the other frames show a clearly older individual). This lines up with what @hukenovs noted in #32 — that some workers may have uploaded photos that weren't of themselves — and it means per-account rather than per-image filtering can miss a minor mixed into an otherwise-adult account.

We're not sharing the full confirmed list here (id's tied to real minors, felt wrong to post in a public thread), but we're happy to send it privately — the full census, user_id-grouped, with our confidence tier per person — if that's useful for re-running your filter or auditing the crowdsourcing pipeline. Just let us know a preferred channel.

Really appreciate the dataset and the consent-oriented design intent — filing this in that spirit, not as a complaint.