En esta página
15.219 14.8.9 Registro de métricas
15.219.1 MF-DATA
asset_integrity_rate, sync_error_ms,
calibration_residual, scorable_fraction,
reference_tier_distribution
15.219.2 MF-POSE
OKS_AP, PCK2D, MPJPE,
PA_MPJPE, PCK3D, visibility_F1,
orientation_error_deg
15.219.3 MF-SCENE
mask_AP, semantic_mIoU, PQ,
SQ, RQ, PartPQ, VPQ,
normal_error_deg, depth_RMSE
15.219.4 MF-TRACK
HOTA, DetA, AssA,
LocA, ID_switches_per_minute,
fragmentation
15.219.5
MF-CONTACT
contact_event_F1, contact_graph_edge_F1,
support_AUPR, onset_MAE_ms,
offset_MAE_ms, dwell_error_ms
15.219.6 MF-LOAD
ordinal_macro_F1, force_MAE_N,
force_RMSE_N, force_MAE_BW,
impulse_error, force_vector_angle_error
15.219.7 MF-PHASE
balanced_frame_accuracy, edit_score,
F1_at_10, F1_at_25, F1_at_50,
boundary_F1, boundary_MAE_ms
15.219.8 MF-KIN
trajectory_RMSE_m, velocity_RMSE_mps,
acceleration_RMSE_mps2, angular_error_deg,
feature_bias, feature_CCC,
DTW_distance
15.219.9 MF-CLASS
macro_F1, micro_F1,
balanced_accuracy, mAP,
hierarchical_F1, Jaccard_multilabel,
contrast_balanced_accuracy
15.219.10
MF-UNCERTAINTY
NLL, Brier,
ECE_reported_with_bins,
classwise_calibration_error, unknown_AUROC,
unknown_AUPR, FPR95, OSCR,
AURC
15.219.11
MF-CONFORMAL
marginal_coverage, group_coverage,
prediction_set_size, singleton_rate,
empty_set_rate
15.219.12
MF-SYSTEM
latency_p50_ms, latency_p95_ms,
latency_p99_ms, throughput_fps,
real_time_factor, peak_RAM_MB,
peak_VRAM_MB, energy_Wh_per_attempt,
crash_rate
15.219.13
MF-HUMAN
time_to_insight_s, decision_accuracy,
coach_agreement, harmful_acceptance_rate,
appropriate_reliance, review_burden,
usability_score
15.219.14 Reglas métricas
- Reportar numerador, denominador, exclusiones y soporte.
- Separar macro, micro, weighted y per-class cuando correspondan.
- Usar precision–recall en clases raras además de ROC; PR refleja mejor el valor predictivo bajo desbalance [P48-R16].
- Brier y NLL evalúan probabilidad; accuracy no evalúa calibración [P48-R17, P46-R14].
- ECE debe declarar bins y complementarse con análisis por clase/grupo; no es una propiedad única del modelo [P48-R14].
- AURC resume el intercambio riesgo–cobertura para sistemas que pueden abstenerse [P48-R6, P46-R25].
- Conformal prediction debe reportar cobertura y eficiencia del set, y no prometer cobertura condicional universal [P48-R15, P46-R15].