15.224 14.8.14 Robustez y stress testing
| ID | Eje | Niveles | Propiedad a conservar |
|---|---|---|---|
ROB-01 |
illumination | exposure shift and local shadow | calibration and unknown behavior |
ROB-02 |
whiteBalance | temperature/tint shift | route identity not color-only |
ROB-03 |
motionBlur | controlled kernel or measured fast motion | boundary uncertainty |
ROB-04 |
defocus | optical or synthetic blur | quality gate |
ROB-05 |
compression | bitrate and codec variation | temporal identity |
ROB-06 |
resolution | downsample profiles | metric coordinate provenance |
ROB-07 |
frameRate | subsampling with real timestamps | derivative validity |
ROB-08 |
frameDrop | random and burst loss | missingness mask |
ROB-09 |
timestampJitter | controlled nonuniform timing | time-based derivatives |
ROB-10 |
cameraShake | global motion | world/scene stabilization |
ROB-11 |
viewpoint | azimuth/elevation/distance shift | 3D and occlusion uncertainty |
ROB-12 |
crop | body/object out-of-frame | out-of-frame state |
ROB-13 |
personOcclusion | limb/torso occlusion | unknown rather than hallucination |
ROB-14 |
holdOcclusion | athlete covers scene object | static scene identity |
ROB-15 |
chalkAndWear | appearance change | geometry over color |
ROB-16 |
backgroundDistractor | people and moving objects | athlete identity |
ROB-17 |
depthNoise | invalid pixels and edge bleed | invalid-is-missing |
ROB-18 |
calibrationDrift | intrinsic/extrinsic perturbation | residual detection and block |
ROB-19 |
forceSensorBias | offset/gain/drift | sensor QA and uncertainty |
ROB-20 |
sensorDropout | one or multiple modalities absent | declared lower operating mode |
ROB-21 |
wallAngle | slab through roof | group reporting |
ROB-22 |
athleteAnthropometry | height/reach/body proportions strata | normalized and native outputs |
ROB-23 |
movementNovelty | held-out class and novel composition | open-set rejection |
ROB-24 |
longAttempt | duration and repeated similar episodes | memory and identity consistency |
Las corrupciones se aplican con severidad controlada y siempre se conserva el resultado clean correspondiente. Los benchmarks de corrupciones muestran la necesidad de medir degradación fuera de los datos limpios [P48-R5].
15.224.1 Perturbación válida
Una perturbación debe: - preservar la etiqueta o declarar cuándo deja de preservarla; - registrar parámetros y seed; - no modificar accidentalmente timestamps, calibración o compresión fuera del eje estudiado; - compararse pareadamente contra el mismo intento limpio; - medir tanto error como abstención segura.