Evidence for 182.C.19.03: track identity through profile turns, partial
occlusion, camera motion, lighting changes, cuts, interactions, and multiple
people.
Two failures, kept apart#
Association fails when the tracker loses the face. It produces fragmentation — more tracks than people. That is the safe failure: a broken track loses continuity and whoever picks it up starts again.
Identity fails when a track continues and the person changes. That is a swap, and it is the dangerous one: the take reports the wrong person with full confidence.
Measuring them apart is the whole value, and the measurement shows they fail under different conditions.
Method#
Six 24-frame sequences built from the committed frame
../2026-09-10-openrouter-face-detection/two-foreground-subjects-plus-bystanders.png
by real geometry and photometry — a moving crop window is a camera move, a black
rectangle is an occluder, an exposure ramp is a lighting change, a splice is a
cut, and two real face crops moving past each other is an interaction between
two people whose pixels are genuinely theirs. Nothing paints a synthetic face.
Detection with YuNet, identity with SFace (128-d cosine), association by IoU alone — deliberately, because mixing the embedding into the association would make the tracker paper over the identity failure being measured.
Two subjects were enrolled: the faces above 0.10 relative frame height. They cross-score at 0.230, so they are genuinely different people.
Results#
| Sequence | tracks | confident tracks | swaps | min cosine | tracks matching nobody |
|---|---|---|---|---|---|
| camera-pan | 49 | 2 | 0 | 0.889 | 47 |
| camera-zoom | 9 | 2 | 0 | 0.901 | 7 |
| lighting-ramp | 6 | 2 | 0 | 0.643 | 4 |
| cut | 9 | 4 | 0 | 0.945 | 5 |
| crossing | 3 | 3 | 0 | 0.388 | 0 |
| occlusion | 6 | 2 | 2 | 0.099 | 4 |
- Camera motion: identity holds (0 swaps, 0.889 minimum) while association fragments into 49 tracks — a pan moves a small face further than its own width between frames and IoU cannot bridge it. One failure, not both.
- Lighting: holds through a −1 to +1 stop ramp, and costs 0.25 of cosine (0.643 against 0.889–0.945 elsewhere). A threshold fitted under even lighting will be tighter than a lit scene can support.
- Cut: correctly breaks the tracks — 4 for 2 people, 0 swaps — which is what a cut should do rather than continuing one across it.
- Interaction: 0 swaps through the overlap; one track split, which is the safe failure.
- Occlusion is the one that breaks. Two identity swaps inside the foreground tracks and a minimum cosine of 0.099: as the occluder covers the face the embedding drifts until the nearest enrolled subject is the other person, and the track carries on with them. Identity assignment is suppressed under occlusion rather than trusted.
The finding that governs every assignment#
47 of the 49 pan tracks matched nobody — their highest mean cosine against the two enrolled subjects was 0.243 — and every one of them was still assigned a subject name, because nearest-neighbour always has a nearest.
An identity assignment with no bound calibration names somebody for every face
in frame. That is the same conclusion 182.C.07.03 reached about pack identity
and it takes the same remedy: return unassigned, and let 182.C.19.08 fit the
cutoff on the 182.C.33 corpus.
The enrolment floor#
Enrolling all six detected faces instead of the two above 0.10 relative height
changed the swap counts in both directions — occlusion 2 → 5, camera-zoom 4
→ 2 — because four unusable bystander embeddings act as absorbers for ambiguous
frames. An enrolment below the measurability floor is a phantom identity, and
assertEnrollable refuses it.
Not tested#
Profile turns. Rotating a face away from camera is a 3D transformation of a
head, and a still gives no geometry to rotate; a sheared or perspective-warped
frontal face is a warped frontal face, and a tracker that survived it would have
proved nothing. Needs recorded footage — 182.C.33. assertConditionTested
refuses to claim it was handled.
Reproducing#
python tools/human-video/evaluator/make_tracking_sequences.py --out /tmp/tracking
python tools/human-video/evaluator/identity_tracking.py \
--enroll /tmp/tracking/enroll.png --sequence /tmp/tracking/occlusion-ramp
Embeddings are biometric data and are never committed — only the cosines
computed from them, following tools/human-video/face-embedding.py.
Per-item status this evidence supports#
| Item | Was | Is |
|---|---|---|
182.C.19.03 |
not-started |
done for six of seven conditions; profile turns need 182.C.33 footage |
182.C.19.08 |
pending | a third reason thresholds cannot be universal: lighting alone cost 0.25 of cosine on one ramp |