14. Señales para visión artificial — 15.233 14.8.23 Atribución de errores

Inicio · Capítulos · 14. Señales para visión artificial

15.233 14.8.23 Atribución de errores

ID	Etapa	Ejemplo
`ERR-DATA`	capture/dataset	missing frames, wrong sync or corrupted asset
`ERR-CAL`	calibration	intrinsic/extrinsic drift
`ERR-POSE-LOC`	pose	keypoint localization error
`ERR-POSE-ID`	pose/tracking	athlete or limb identity swap
`ERR-SCENE-CLS`	scene	hold/volume class confusion
`ERR-SCENE-GEO`	scene	surface normal or depth invalid
`ERR-CONTACT-FP`	contact	proximity interpreted as contact
`ERR-CONTACT-FN`	contact	occluded functional contact missed
`ERR-LOAD`	load	touch interpreted as primary support
`ERR-BOUNDARY`	phase	early/late boundary or oversegmentation
`ERR-LANE`	phase	simultaneous actions collapsed into one label
`ERR-DERIV`	kinematics	noise amplification or invalid interpolation
`ERR-CLASS-CONTRAST`	classification	confusion within priority contrast set
`ERR-CLASS-HIER`	classification	inconsistent family and leaf labels
`ERR-UNKNOWN`	open-set	novel action forced to known label
`ERR-CALIBRATION`	uncertainty	high confidence in wrong prediction
`ERR-INTEGRATION`	system	schema/frame mismatch across components
`ERR-HUMAN`	human interaction	misleading evidence causes inappropriate reliance

15.233.1 Matriz de propagación

pose/scene error
    -> wrong contact candidate
        -> wrong phase boundary
            -> distorted kinematics
                -> wrong movement label
                    -> misleading coach evidence

El benchmark guarda outputs intermedios para diferenciar un clasificador incorrecto de un clasificador que recibió contactos equivocados.