15.214 14.8.4 Tareas oficiales
| ID | Tarea | Unidad | Referencia | Métricas principales |
|---|---|---|---|---|
TASK-01 |
Pose2DAndVisibility | person-frame/keypoint | adjudicated 2D keypoints and visibility | OKS-AP PCK2D visibility-F1 out-of-frame detection |
TASK-02 |
Pose3DAndBodyFrames | person-frame/joint | multiview or motion-capture 3D | MPJPE PA-MPJPE PCK3D bone-length error frame orientation error |
TASK-03 |
ScenePanopticPartsAndGeometry | frame/object/patch | adjudicated masks, instances, parts and surveyed geometry | PQ PartPQ mask AP surface normal error depth error |
TASK-04 |
AthleteAndObjectTracking | track/frame | persistent identities | HOTA DetA AssA LocA ID switches |
TASK-05 |
ContactDetection | body-region/object/time | adjudicated contact episodes | event precision/recall/F1 edge F1 onset/offset error dwell error |
TASK-06 |
FunctionalSupportAndLoadState | contact episode | instrumented or expert-adjudicated state | ordinal macro F1 support AUCPR state transition error interval coverage |
TASK-07 |
MeasuredOrInferredForce | instrumented contact/sample | force sensor synchronized trace | MAE N RMSE N MAE bodyweight-normalized impulse error vector angle error |
TASK-08 |
PhaseBoundaryDetection | boundary event | uncertain adjudicated interval | boundary F1 by tolerance boundary MAE ms interval hit rate early/late bias |
TASK-09 |
PhaseSegmentation | attempt timeline | hierarchical multilanes | balanced frame accuracy edit score segmental F1@10/25/50 lane consistency |
TASK-10 |
KinematicFeatureReconstruction | trajectory/phase summary | validated biomechanical trajectory | position RMSE velocity RMSE angular error feature bias agreement limits |
TASK-11 |
MovementMultiLabelClassification | movement episode | adjudicated canonical labels | macro F1 micro F1 mAP hierarchical F1 label ranking loss |
TASK-12 |
FailurePatternDetection | failure episode | expert-adjudicated failure labels | macro F1 event AP time-to-detection false alert rate |
TASK-13 |
ContrastSetDifferentiation | paired/contrast episode | high-priority contrast label | pairwise balanced accuracy contrast margin abstention-aware accuracy evidence completeness |
TASK-14 |
OpenSetAndSelectivePrediction | episode | known/unknown/insufficient state | AUROC unknown AUPR unknown FPR95 OSCR AURC risk at target coverage |
TASK-15 |
CalibrationAndPredictionSets | episode/class | empirical correctness | NLL Brier classwise calibration coverage prediction set size |
TASK-16 |
EndToEndAttemptReconstruction | complete attempt | all adjudicated layers | attempt graph score hold sequence edit contact-phase consistency semantic reconstruction score |
TASK-17 |
OperationalInference | device/run | benchmark harness | latency p50/p95/p99 throughput real-time factor peak memory energy per attempt |
TASK-18 |
CoachDecisionSupport | coach-case pair | prespecified coaching task and outcome | time to insight decision agreement error detection harmful acceptance calibrated trust |
Las tareas de pose usan medidas distintas para localización 2D, 3D,
visibilidad y orientación. PA-MPJPE incluye alineamiento y
no debe sustituir MPJPE; una mejora después de alineamiento
puede ocultar error global de escala o posición. COCO usa OKS/AP para
keypoints y trabajos recientes extienden la evaluación a puntos fuera de
imagen [P41-R13, P48-R11].
Para escena, PQ combina reconocimiento y calidad de
segmentación; PartPQ añade partes y VPQ
extiende coherencia temporal [P42-R2, P48-R12, P48-R3]. Para tracking,
HOTA separa detección, asociación y localización [P48-R2].
La segmentación temporal debe incluir edit score y F1 por solapamiento además de accuracy por frame, ya que una secuencia sobresegmentada puede conservar muchos frames correctos [P44-R5].