Inicio · Capítulos · 14. Señales para visión artificial

15.193 14.7.19 Particiones y prevención de fuga

15.193.1 Orden conservador de agrupación

athleteId → problemId → captureSessionId → gymId → sceneRevisionId → calendarPeriod → deviceConfigurationId

15.193.2 Reglas

• Assign raw capture groups before deriving clips or frames.
• All overlapping windows from one attempt remain in one split.
• A route reset or moved hold creates a new scene revision, but athlete/session grouping can still forbid cross-split use.
• Test labels and adjudication rationales remain quarantined from model development.
• Open-set evaluation declares held-out labels, held-out attributes and held-out sites separately.
• Calibration captures may be shared only when they contain no athlete/action examples and this is documented.
• Every split stores deterministic assignments and a leakage audit, not just percentages.

15.193.3 Vistas de evaluación

• Within-distribution grouped test
• Unseen-athlete test
• Unseen-problem test
• Unseen-gym/site test
• Unseen-camera/view test
• Temporal holdout
• Held-out-class open-set test
• Cross-modality degradation test

15.193.4 Prohibido

• Prohibido: Random frame split.
• Prohibido: Random clip split before grouping.
• Prohibido: Same attempt in multiple splits.
• Prohibido: Near-duplicate re-encoding across splits.
• Prohibido: Using test disagreement to rewrite training labels.

capture group -> immutable split -> clips/windows/frames
NOT: frames -> random split -> discover leakage later