FR-CULL-8a — Per-face eye state and head pose, as evidence #56

New Issue

2026-09-19T10:20:52Z

dtourolle commented

2026-09-19 10:20:52 +00:00

For every detected face, record one open-probability per eye and a yaw/pitch/roll, as derived data under FR-CULL-12. Added to the spec 2026-09-19 (7db999c); no code exists.

Why

"Everyone's eyes open" and "facing the camera" are the two questions a group or event cull asks of every frame, and both are facts the detector has already half paid for: eye crops fall out of the five landmarks, and head pose is geometry against a generic template with no model at all. Gaze is deferred (§7) because no weights with a clean data chain exist; pose is the proxy.

Deliverable

Eye state from OCEC (MIT code and weights, ODC-By/Apache data — the first clean chain found; D13), a 24×40 crop per eye cut from the landmarks, run in the same indexing pass as detection. Its published F1 is on its own crops; ours are measured on this library before being believed.
Head pose solved from the five landmarks against a generic template; "facing the camera" is a stated band.
Both stored in the catalog only, recomputed by re-indexing, never in a sidecar.
Both terms in the selector language, composable with a person and every other term (FR-CULL-11).
Both feed FR-CULL-13's evidence; neither may set a rating or a flag.

Acceptance

On ≥500 labelled faces from the reference library, eye state within a stated tolerance of published F1, reported separately for glasses, profile, and faces under 60 px
Facing-the-camera agrees with a hand-labelled split at a stated rate
Indexing stays inside FR-CULL-8's acceptance (grid never below NFR-P9)
TRACES: FR-CULL-8a on the classifier, the pose solver and the schema columns

**For every detected face, record one open-probability per eye and a yaw/pitch/roll, as derived data under FR-CULL-12.** Added to the spec 2026-09-19 (7db999c); no code exists. ## Why "Everyone's eyes open" and "facing the camera" are the two questions a group or event cull asks of every frame, and both are facts the detector has already half paid for: eye crops fall out of the five landmarks, and head pose is geometry against a generic template with no model at all. Gaze is deferred (§7) because no weights with a clean data chain exist; pose is the proxy. ## Deliverable - Eye state from OCEC (MIT code and weights, ODC-By/Apache data — the first clean chain found; D13), a 24×40 crop per eye cut from the landmarks, run in the same indexing pass as detection. Its published F1 is on its own crops; ours are measured on this library before being believed. - Head pose solved from the five landmarks against a generic template; "facing the camera" is a stated band. - Both stored in the catalog only, recomputed by re-indexing, never in a sidecar. - Both terms in the selector language, composable with a person and every other term (FR-CULL-11). - Both feed FR-CULL-13's evidence; neither may set a rating or a flag. ## Acceptance - [ ] On ≥500 labelled faces from the reference library, eye state within a stated tolerance of published F1, reported separately for glasses, profile, and faces under 60 px - [ ] Facing-the-camera agrees with a hand-labelled split at a stated rate - [ ] Indexing stays inside FR-CULL-8's acceptance (grid never below NFR-P9) - [ ] `TRACES: FR-CULL-8a` on the classifier, the pose solver and the schema columns

dtourolle added the catalog unmet-requirement size:M culling labels 2026-09-19 10:20:52 +00:00

dtourolle added a new dependency 2026-09-19 10:21:01 +00:00

#55 FR-CULL-13 — Evidence never becomes a verdict, and a test proves it

dtourolle commented

2026-09-24 11:18:39 +00:00

Progress check (2026-09-24).

eye state end to end: OCEC classifier plus sunglasses (core/dr-face/src/classify.rs, eyes.rs), seven numbers per face in catalog columns and shards (6b51726, b908d86, 54b543f, facb44c), models shipped (6aae4c3), a selector term for "eyes open" (83f4253), a covering index (d706c12)
head pose: nothing outside dr-pano, no "facing the camera" term
the measurement on ≥500 labelled faces split by glasses / profile / <60 px — faces.md §17.4 has only a 25–60 face proxy sample, and M11–M13 are still to measure
indexing throughput against FR-CULL-8's acceptance: none found

Progress check (2026-09-24). - [x] eye state end to end: OCEC classifier plus sunglasses (`core/dr-face/src/classify.rs`, `eyes.rs`), seven numbers per face in catalog columns and shards (6b51726, b908d86, 54b543f, facb44c), models shipped (6aae4c3), a selector term for "eyes open" (83f4253), a covering index (d706c12) - [ ] head pose: nothing outside dr-pano, no "facing the camera" term - [ ] the measurement on ≥500 labelled faces split by glasses / profile / <60 px — faces.md §17.4 has only a 25–60 face proxy sample, and M11–M13 are still to measure - [ ] indexing throughput against FR-CULL-8's acceptance: none found

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Blocks

#55 FR-CULL-13 — Evidence never becomes a verdict, and a test proves it

dtourolle/DarkRoom

Reference: dtourolle/DarkRoom#56