Enciclopedia Boulder v1.0

Inicio · Capítulos · 14. Señales para visión artificial

15.226 14.8.16 Protocolo estadístico

Elemento Regla
estimandFirst Cada tarea declara población objetivo, unidad de análisis, score y dirección antes de ejecutar test.
primaryUnit Attempt unless a task requires session, athlete, track, event or coach-case; frames are not independent replicates.
confidenceIntervals Cluster bootstrap or another prespecified dependence-aware method; default reporting level 95%, never point estimate alone.
pairedComparison Paired cluster bootstrap, permutation or justified paired test on identical test units.
stochasticModels Repeat prespecified seeds; report within-seed and between-seed variation separately.
multipleComparisons Declare primary hypotheses and control multiplicity for families of confirmatory comparisons.
effectSize Report absolute and relative difference plus confidence interval; p-value alone is insufficient.
minimumSupport Defined prospectively per class/group and claim; insufficient support yields NotEstimable rather than zero.
missingResults Report excluded, unscorable, timeout and failed attempts with reasons and sensitivity analysis.
worstGroup Report group support and interval; do not select groups post hoc solely because performance was low.
thresholdSelection All thresholds selected on train/calibration/validation only and frozen before final test.
testReuse Each official test opening is logged; repeated tuning invalidates confirmatory status.

Los frames contiguos del mismo intento son dependientes. El bootstrap debe remuestrear clusters coherentes —por ejemplo intentos o sesiones— y no millones de frames como observaciones independientes [P48-R9].

Las comparaciones de modelos deben usar unidades pareadas y pruebas adecuadas al diseño; elegir un test sólo después de ver los resultados invalida la interpretación confirmatoria [P48-R8, P48-R10].

15.226.1 Reporte mínimo por score

estimate
confidence interval
support / cluster count
reference tier
dataset view
threshold source
exclusions and failure count