En esta página
15.37 14.2.13 Segmentación, refinamiento y tracking
| Rol | Uso |
|---|---|
| Production semantic/instance model | Fine-tuned domain model predicts hold/volume/scene classes and masks. |
| Promptable segmenter | SAM/SAM2 assists annotation, correction and temporal propagation; it does not supply climbing semantics by itself. |
| Depth geometry | Splits touching masks, verifies protrusion, recovers surface patches and rejects implausible regions. |
| Static map projector | Maintains identity during occlusion and supplies amodal expectations. |
| Human review | Resolves route membership, ambiguous objects, topology changes and competition validity. |
SAM permite segmentación mediante prompts y SAM 2 añade memoria temporal para video [P42-R3, P42-R4]. En el pipeline propuesto son herramientas de anotación, refinamiento y propagación. No sustituyen una clase de dominio ni una revisión de topología.
15.37.1 Estados de tracking
Tentative, Confirmed,
Occluded, OutOfFrame, Retired,
Reidentified
15.37.2 Variables de asociación
- projected mask IoU
- 3D centroid distance
- surface/mesh overlap
- parent relationship
- appearance embedding
- route/marker consistency
- shape descriptor
- temporal continuity
15.37.3 Reglas split/merge
- A mask split caused by athlete occlusion keeps one sceneObjectId with multiple visible components.
- Two touching holds require separate IDs when static geometry or annotation resolves the boundary.
- A merge/split changes scene topology only after persistent multi-frame or manual evidence.
- Child holds remain distinct from their parent volume even if their colors match.
Mask quality debe almacenarse separada de la confianza de clase; trabajos como Mask Scoring R-CNN muestran que ambas puntuaciones no están necesariamente alineadas [P42-R13]. Para presas pequeñas, la precisión de borde también debe evaluarse explícitamente [P42-R14].