14. Señales para visión artificial — 15.150 14.6.9 Arquitectura de referencia

Inicio · Capítulos · 14. Señales para visión artificial

15.150 14.6.9 Arquitectura de referencia

ID	Componente	Entrada	Salida
ARCH-01	QualityGate	all streams	validity masks and insufficiency reasons
ARCH-02	EpisodeProposal	14.4 events/phases	variable-length candidate windows
ARCH-03	PoseEncoder	joint/segment series	skeleton embedding with visibility masks
ARCH-04	SceneContactGraphEncoder	body–surface–contact graph	interaction embedding
ARCH-05	KinematicEncoder	14.5 series and aggregates	scale-aware motion embedding
ARCH-06	OptionalAppearanceEncoder	RGB/depth crops	masked visual embedding
ARCH-07	TemporalFusion	multimodal embeddings	causal or bidirectional episode representation
ARCH-08	AttributeHeads	episode representation	label-space probabilities and evidence
ARCH-09	CanonicalLabelHead	attributes + representation	310 movement and 34 failure logits
ARCH-10	ContrastiveDifferentialHead	candidate confusion set	pair/set margins and decisive evidence
ARCH-11	OpenSetAndCalibrationHead	logits/evidence/embedding	calibrated probabilities, novelty and reject state
ARCH-12	OntologyDecoder	all heads	hierarchy-consistent multi-label predictions
ARCH-13	EvidenceLedger	features, constraints and heads	supporting, contradicting and missing evidence
ARCH-14	HumanReviewQueue	selective outputs	review package and adjudication feedback

quality gate -> episode proposal
   -> pose encoder
   -> scene/contact graph encoder
   -> kinematic encoder
   -> optional RGB-D/sensor encoders
        -> temporal fusion
             -> attribute heads
             -> canonical multi-label heads
             -> differential contrast heads
             -> open-set + calibration
                  -> ontology decoder -> evidence ledger -> selective publication

Action segmentation architectures como ASFormer, ASRF y DiffAct ilustran la necesidad de contexto largo, refinamiento y tratamiento explícito de fronteras [P46-R6, P46-R7, P46-R8]. Las fronteras de 14.4 se conservan como intervalos de incertidumbre, no como cortes perfectos.