give segmentation QA the same threshold-independent view as detection

Detection QA reports a precision/recall curve, average precision and a
calibration sweep; segmentation QA reported a single operating point. Both rank
their outputs by confidence, so the same view applies, and the asymmetry meant
the two panels answered different questions about comparable runs — an
inconsistency introduced when detection gained the curve.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backend/app/api/routes/segmentation.py

@@ -223,5 +223,6 @@ def compare_segmentation_run_with_reference(
             iou_threshold=payload.iou_threshold,
             class_name=payload.class_name,
             min_confidence=payload.min_confidence,
+            calibration_thresholds=payload.calibration_thresholds,
         )
     )

backend/app/schemas/segmentation.py

@@ -32,6 +32,8 @@ class SegmentationQaRequest(BaseModel):
     iou_threshold: float = Field(default=0.5, ge=0.0, le=1.0)
     class_name: str | None = None
     min_confidence: float | None = Field(default=None, ge=0.0, le=1.0)
+    # Read off the one matching pass, exactly as for detection.
+    calibration_thresholds: list[float] = Field(default_factory=list, max_length=32)
 
 
 class SegmentationRunResponse(BaseModel):

backend/app/services/segmentation_service.py

@@ -21,6 +21,7 @@ from app.schemas.segmentation import (
     SegmentationRunResponse,
 )
 from app.services.detection_georeferencing import pixel_points_to_epsg4326_polygon
+from app.services.detection_metrics_service import DetectionMetricsService
 from app.services.detection_qa_service import DetectionQaService
 from app.services.detection_service import DetectionService
 from app.services.dataset_consumption_gate_service import DatasetConsumptionGate
@@ -328,6 +329,7 @@ class SegmentationService:
         iou_threshold: float = 0.5,
         class_name: str | None = None,
         min_confidence: float | None = None,
+        calibration_thresholds: list[float] | None = None,
     ) -> dict[str, Any]:
         run = db.get(AnalysisRun, analysis_run_id)
         if not run or run.analysis_type != "segmentation":
@@ -474,6 +476,18 @@ class SegmentationService:
             reference_geometries,
             iou_threshold,
         )
+        # Segmentation confidences rank the same way detections do, so the same
+        # threshold-independent view applies. Detection had it and segmentation
+        # did not, which made the two QA panels answer different questions.
+        precision_recall_curve = DetectionMetricsService.precision_recall_curve(
+            candidate_geometries,
+            reference_geometries,
+            iou_threshold=iou_threshold,
+        )
+        calibration_sweep = DetectionMetricsService.calibration_sweep(
+            precision_recall_curve,
+            thresholds=list(calibration_thresholds or []),
+        )
         mean_iou = None if not evidence.match_iou_values else sum(evidence.match_iou_values) / len(evidence.match_iou_values)
         precision = evidence.matches / (evidence.matches + evidence.false_positives) if evidence.matches + evidence.false_positives > 0 else None
         recall = evidence.matches / (evidence.matches + evidence.false_negatives) if evidence.matches + evidence.false_negatives > 0 else None
@@ -505,6 +519,8 @@ class SegmentationService:
                 "warnings": coverage_warnings + evidence.warnings,
                 "unsupported_geometry": evidence.unsupported,
                 "coverage": coverage_summary,
+                "precision_recall_curve": precision_recall_curve,
+                "calibration_sweep": calibration_sweep,
                 "match_evidence": evidence.match_evidence,
                 "false_positive_evidence": evidence.false_positive_evidence,
                 "false_negative_evidence": evidence.false_negative_evidence,
@@ -516,6 +532,9 @@ class SegmentationService:
                 "mean_iou": mean_iou,
                 "false_positive_count": evidence.false_positives,
                 "false_negative_count": evidence.false_negatives,
+                "average_precision": precision_recall_curve["average_precision"],
+                "best_f1": precision_recall_curve["best_f1"],
+                "best_f1_threshold": precision_recall_curve["best_f1_threshold"],
             },
         )
         return {
@@ -537,6 +556,8 @@ class SegmentationService:
             "iou_threshold": iou_threshold,
             "warnings": coverage_warnings + evidence.warnings,
             "coverage": coverage_summary,
+            "precision_recall_curve": precision_recall_curve,
+            "calibration_sweep": calibration_sweep,
             "match_evidence": evidence.match_evidence,
             "false_positive_evidence": evidence.false_positive_evidence,
             "false_negative_evidence": evidence.false_negative_evidence,

backend/tests/test_sprint9_segmentation_foundation.py

@@ -362,6 +362,11 @@ def test_segmentation_qa_persists_quality_check_and_metrics() -> None:
         "mean_iou",
         "false_positive_count",
         "false_negative_count",
+        # Threshold-independent, as detection QA already reported. The two
+        # panels previously answered different questions about the same run.
+        "average_precision",
+        "best_f1",
+        "best_f1_threshold",
     ]
 
 

docs/API_CONTRACTS.md

@@ -1981,7 +1981,12 @@ compare two models whose calibration differs; the curve can. `average_precision`
 existing ones.
 
 Segmentation QA (`POST /api/v1/segmentation/runs/{analysis_run_id}/qa/reference`)
-applies the same tile-coverage clipping and returns the same `coverage` block.
+applies the same tile-coverage clipping and returns the same `coverage` block,
+and now also the same `precision_recall_curve`, `calibration_thresholds` /
+`calibration_sweep` and average-precision metrics. Segmentation confidences rank
+the same way detection confidences do, so the same threshold-independent view
+applies; without it the two QA panels answered different questions about
+comparable runs.
 Without it, every reference feature outside the inferred tiles counted as a
 false negative and recall was understated by an arbitrary amount.
 

frontend/src/types.ts

@@ -1620,6 +1620,8 @@ export interface SegmentationQaRequest {
   iou_threshold: number
   class_name?: string | null
   min_confidence?: number | null
+  /** Read off the one matching pass, exactly as for detection. */
+  calibration_thresholds?: number[]
 }
 
 export type SegmentationQaResult = DetectionQaResult