Target related hard-negative contexts

scripts/build_failure_driven_yolo_sampling.py

@@ -12,6 +12,15 @@ from pathlib import Path
 from typing import Any
 
 
+PRECISION_NEGATIVE_CONTEXTS = {
+    "coastal-urban": {"port-hard-negative", "dunes-negative"},
+    "industrial": {"industrial-hard-negative", "rail-hard-negative", "port-hard-negative"},
+    "ribbon-development": {"farmland-hard-negative", "forest-hard-negative"},
+    "rural-town": {"farmland-hard-negative", "forest-hard-negative", "quarry-hard-negative"},
+    "regional-architecture": {"forest-hard-negative", "quarry-hard-negative"},
+}
+
+
 def file_sha256(path: Path) -> str:
     digest = hashlib.sha256()
     with path.open("rb") as stream:
@@ -94,6 +103,12 @@ def build_sampling(
             and metrics["precision"] < gates["min_region_precision"]
         ):
             weak_precision_contexts.add(key)
+    targeted_negative_contexts = set(weak_precision_contexts)
+    for region, context in weak_precision_contexts:
+        targeted_negative_contexts.update(
+            (region, related)
+            for related in PRECISION_NEGATIVE_CONTEXTS.get(context, set())
+        )
 
     base_paths_by_region: dict[str, list[str]] = {}
     extra_paths_by_region: dict[str, list[str]] = {}
@@ -120,12 +135,19 @@ def build_sampling(
         if tile["label_count"] == 0 and (background_failed or region in weak_precision_regions):
             repeat = (
                 context_negative_repeat
-                if context_key in weak_precision_contexts
+                if context_key in targeted_negative_contexts
                 else negative_repeat
             )
         path = str(Path(tile["image_path"]).resolve())
         base_paths_by_region.setdefault(region, []).append(path)
-        extra_paths_by_region.setdefault(region, []).extend([path] * (repeat - 1))
+        extras = extra_paths_by_region.setdefault(region, [])
+        repeated = [path] * (repeat - 1)
+        if tile["label_count"] == 0 and context_key in targeted_negative_contexts:
+            # Preserve the most diagnostic hard-negative repeats when the
+            # regional cap has to remove lower-priority repetition.
+            extras[:0] = repeated
+        else:
+            extras.extend(repeated)
         selected_samples.add(tile["sample_slug"])
 
     pre_cap_counts = Counter({
@@ -163,6 +185,9 @@ def build_sampling(
         "weak_precision_regions": sorted(weak_precision_regions),
         "weak_recall_contexts": [f"{region}:{context}" for region, context in sorted(weak_recall_contexts)],
         "weak_precision_contexts": [f"{region}:{context}" for region, context in sorted(weak_precision_contexts)],
+        "targeted_negative_contexts": [
+            f"{region}:{context}" for region, context in sorted(targeted_negative_contexts)
+        ],
         "background_gate_failed": background_failed,
         "positive_repeat": positive_repeat,
         "negative_repeat": negative_repeat,