| """Align extracted entities with synthgen ground-truth records (T020/T022). |
| |
| synthgen writes entities-vN.json per revision: {uid, kind, semantic, bbox, |
| text}. Extraction produces one DrawingEntity per synthetic primitive, so a |
| deterministic one-to-one nearest-center assignment within the same structural |
| kind recovers labels (semantic → classifier targets) and cross-revision |
| correspondence (shared uid → matcher positives). |
| """ |
|
|
| from __future__ import annotations |
|
|
| import json |
| from pathlib import Path |
| from typing import Any |
|
|
| from cadmorph.models import DrawingGraph |
|
|
| ALIGN_TOL_REL = 0.02 |
|
|
|
|
| def load_records(path: str | Path) -> list[dict[str, Any]]: |
| return json.loads(Path(path).read_text(encoding="utf-8")) |
|
|
|
|
| def _center(bbox: list[float] | tuple[float, ...]) -> tuple[float, float]: |
| return ((bbox[0] + bbox[2]) / 2.0, (bbox[1] + bbox[3]) / 2.0) |
|
|
|
|
| def align_records(graph: DrawingGraph, records: list[dict[str, Any]]) -> dict[str, dict[str, Any]]: |
| """entity_id -> ground-truth record, one-to-one, globally nearest-first. |
| |
| Only same-kind pairs within ALIGN_TOL_REL of each other are eligible; |
| entities/records without a counterpart stay unassigned. |
| """ |
| tol = graph.revision.sheet_diagonal * ALIGN_TOL_REL |
| candidates: list[tuple[float, str, int]] = [] |
| for entity in graph.entities: |
| ex, ey = _center(entity.bbox) |
| for i, record in enumerate(records): |
| if record["kind"] != entity.kind: |
| continue |
| rx, ry = _center(record["bbox"]) |
| dist = ((ex - rx) ** 2 + (ey - ry) ** 2) ** 0.5 |
| if dist <= tol: |
| candidates.append((dist, entity.entity_id, i)) |
|
|
| assignment: dict[str, dict[str, Any]] = {} |
| used_records: set[int] = set() |
| for dist, entity_id, i in sorted(candidates, key=lambda c: (c[0], c[1], c[2])): |
| if entity_id in assignment or i in used_records: |
| continue |
| assignment[entity_id] = records[i] |
| used_records.add(i) |
| return assignment |
|
|
