Expand YOLO training AOIs safely

2026-07-12 23:42:29 +02:00
parent 53cd38a5b2
commit 0f49c980ba
11 changed files with 310 additions and 7 deletions
@@ -6974,3 +6974,28 @@ Open:
 ## Next recommended pass

 - Train one inactive candidate from the filtered AOI1024 cleanpx dataset, then run the existing positive-AOI matrix and split-background promotion workflow before considering default activation.
+# Sprint 171 - Positive AOI expansion and split safety
+
+## What changed
+
+- Converted the Sprint 170 false-negative evidence into a guarded data action instead of another blind training run.
+- Added Olen, Lille, Oud-Turnhout and Kasterlee center as explicit real-reference training AOIs.
+- Kept Turnhout, Retie, Westerlo and Arendonk-heide as the documented validation holdouts.
+- Added generated `recommended_split` provenance and tile-export validation that rejects unknown samples and manifest-backed holdout leakage.
+- Hardened false-negative portfolio comparison to require identical reference feature identities, not only matching AOI names.
+
+## Local validation
+
+- RED tests proved the expansion/split constants and validation guard were absent before implementation.
+- `python -m pytest backend/tests/test_sprint131_operator_sample_expansion.py backend/tests/test_sprint130_operator_yolo_tile_dataset.py backend/tests/test_sprint156_background_corpus_classification.py backend/tests/test_sprint170_detection_false_negative_audit.py -q`: 23 passed.
+- Official GRB OGC API probes returned building features at all four new AOI centers.
+- Local sample generation was attempted but correctly stopped because the workstation Python lacks the existing GIS runtime extras; the all-in-one Tower runtime is the supported execution environment.
+
+## Known limitations
+
+- No new model has been trained or activated.
+- The four AOIs still require full orthophoto/reference preparation, tile export, structural audit and visual contact-sheet review on Tower.
+
+## Next recommended pass
+
+- Redeploy Tower, refresh the existing AOI1024 manifest so only missing AOIs are downloaded, then compare a `min-label-px=4` expanded dataset against the rejected `min-label-px=12` baseline before deciding whether training is justified.