15.192 14.7.18 Muestreo y cobertura
- athleteId and athlete-independent generalization.
- age band with privacy-preserving reporting.
- experience/skill band with declared measurement.
- body-size and reach bands without identity inference.
- gym/site and route-setter.
- problem/route identity and reset revision.
- wall inclination and 3D geometry.
- hold/volume morphology and route color conditions.
- movement family and 344-label long tail.
- priority contrast set and hard-negative type.
- intentional movement vs failure pattern.
- static, controlled dynamic, flight and coordination regimes.
- camera model, lens, view and distance.
- single-view vs multiview and modality availability.
- lighting, occlusion, clothing and chalk conditions.
- training vs competition context.
- successful vs failed attempt and censored outcome.
- time period for temporal drift.
- unknown/novel actions and non-climbing activity.
- annotation difficulty and disagreement stratum.
15.192.1 Long tail
- Medir ejemplos por grupo independiente, no clips derivados.
- Mantener clases raras aunque no alcancen entrenamiento supervisado; pueden usarse como evaluación cualitativa, few-shot u unknown-near-known.
- Coleccionar hard negatives para los 28 contrast sets de 14.6.
- No equilibrar duplicando el mismo intento o generando ventanas casi idénticas.
- Publicar celdas vacías y claims limitados en lugar de imputar representatividad.
BDD100K ilustra un dataset multitarea con diversidad de condiciones [P47-R18]; COCO, ActivityNet y HACS muestran contratos diferentes para instancias, clips y segmentos [P47-R15, P47-R16, P47-R17]. Ninguno cubre por sí solo la combinación de escena, contacto, fases y movimiento requerida aquí.