From 8155acf969e548cb70bc419bd65fdd06d7856c14 Mon Sep 17 00:00:00 2001
From: Jens <jens@localhost>
Date: Mon, 27 Jul 2026 09:13:28 +0200
Subject: [PATCH] Evaluate regional detector ensembles

---
 .../evaluate_belgium_building_candidate.py    | 53 ++++++++++++++++++-
 1 file changed, 52 insertions(+), 1 deletion(-)

diff --git a/scripts/evaluate_belgium_building_candidate.py b/scripts/evaluate_belgium_building_candidate.py
index 31a4dc57..d4a91790 100644
--- a/scripts/evaluate_belgium_building_candidate.py
+++ b/scripts/evaluate_belgium_building_candidate.py
@@ -52,6 +52,31 @@ def metrics(tp: int, fp: int, fn: int) -> dict[str, float | int]:
     }
 
 
+def ensemble_predictions(
+    primary: list[tuple[tuple[float, float, float, float], float]],
+    secondary: list[tuple[tuple[float, float, float, float], float]],
+    *,
+    match_iou: float,
+    mode: str,
+) -> list[tuple[tuple[float, float, float, float], float]]:
+    unmatched = set(range(len(secondary)))
+    combined = []
+    unmatched_primary = []
+    for box, score in primary:
+        candidates = [(iou(box, secondary[index][0]), index) for index in unmatched]
+        overlap, index = max(candidates, default=(0.0, -1))
+        if overlap < match_iou:
+            unmatched_primary.append((box, score))
+            continue
+        other_box, other_score = secondary[index]
+        unmatched.remove(index)
+        combined.append((tuple((left + right) / 2 for left, right in zip(box, other_box)), (score + other_score) / 2))
+    if mode == "union":
+        combined.extend(unmatched_primary)
+        combined.extend(secondary[index] for index in unmatched)
+    return combined
+
+
 def scale_box(
     box: tuple[float, float, float, float], factor: float, offset_x: float = 0.0, offset_y: float = 0.0
 ) -> tuple[float, float, float, float]:
@@ -106,6 +131,9 @@ def main() -> int:
     parser.add_argument("--box-scale", type=float, default=1.0)
     parser.add_argument("--box-offset-x", type=float, default=0.0)
     parser.add_argument("--box-offset-y", type=float, default=0.0)
+    parser.add_argument("--additional-model", type=Path)
+    parser.add_argument("--ensemble-mode", choices=("consensus", "union"), default="consensus")
+    parser.add_argument("--ensemble-match-iou", type=float, default=0.3)
     parser.add_argument("--proposal-classifier", type=Path)
     parser.add_argument("--proposal-classifier-threshold", type=float, default=0.5)
     parser.add_argument("--proposal-crop-scale", type=float, default=1.4)
@@ -147,8 +175,14 @@ def main() -> int:
         max_det=args.max_det,
         verbose=False,
     )
+    additional_results = None
+    if args.additional_model:
+        additional_results = YOLO(str(args.additional_model)).predict(
+            image_paths, conf=min(args.thresholds), device=args.device, augment=args.augment,
+            imgsz=args.imgsz, max_det=args.max_det, verbose=False,
+        )
     observations: list[dict[str, Any]] = []
-    for tile, result in zip(tiles, results, strict=True):
+    for result_index, (tile, result) in enumerate(zip(tiles, results, strict=True)):
         height, width = result.orig_shape
         predictions = [
             (
@@ -159,6 +193,20 @@ def main() -> int:
             )
             for box, score in zip(result.boxes.xyxy.cpu().tolist(), result.boxes.conf.cpu().tolist(), strict=True)
         ]
+        if additional_results is not None:
+            secondary = [
+                (
+                    scale_box(tuple(map(float, box)), args.box_scale, args.box_offset_x, args.box_offset_y),
+                    float(score),
+                )
+                for box, score in zip(
+                    additional_results[result_index].boxes.xyxy.cpu().tolist(),
+                    additional_results[result_index].boxes.conf.cpu().tolist(), strict=True,
+                )
+            ]
+            predictions = ensemble_predictions(
+                predictions, secondary, match_iou=args.ensemble_match_iou, mode=args.ensemble_mode
+            )
         if proposal_classifier is not None:
             from PIL import Image
             import torch
@@ -229,6 +277,9 @@ def main() -> int:
         "box_scale": args.box_scale,
         "box_offset_x": args.box_offset_x,
         "box_offset_y": args.box_offset_y,
+        "additional_model": str(args.additional_model) if args.additional_model else None,
+        "ensemble_mode": args.ensemble_mode if args.additional_model else None,
+        "ensemble_match_iou": args.ensemble_match_iou if args.additional_model else None,
         "proposal_classifier": str(args.proposal_classifier) if args.proposal_classifier else None,
         "proposal_classifier_threshold": args.proposal_classifier_threshold if args.proposal_classifier else None,
         "proposal_crop_scale": args.proposal_crop_scale if args.proposal_classifier else None,