Promote expanded building detector profile

2026-07-13 02:21:42 +02:00
parent d02b4e5219
commit 1c16313ae6
7 changed files with 71 additions and 31 deletions
@@ -7039,3 +7039,25 @@ Open:
 - The expanded live export exposed that Arendonk-heide remained configured as a holdout while all of its low-variance tiles were correctly filtered out.
 - Added `retained_validation_sample_slugs` and `empty_validation_sample_slugs` to tile dataset summaries so configured and actual validation coverage cannot be confused.
 - Added a focused regression test and kept filtering behavior unchanged; no blank tile was reintroduced.
+
+# Sprint 173 - Expanded building model promotion
+
+## Runtime evidence
+
+- Completed inactive 50-epoch CPU training for `geointel-building-yolov8s-aoi1024expandedminpx4vis035e50.pt` from the expanded 20-source real-data corpus.
+- Trained-model SHA256: `a8a79cf5b0bdc19a0245acc322cf77232c335e222bd5f3c00a17d5f29402c196`.
+- Training summary recorded 45,892 labels across 171 retained tiles; final training metrics were precision `0.55558`, recall `0.35206`, mAP50 `0.27440` and mAP50-95 `0.10107`.
+- The persisted seven-AOI positive matrix recommended tile size `512`, overlap `64` and threshold `0.15`: mean precision `0.6470590036`, recall `0.4699913837`, F1 `0.5432865391`, and minimum per-AOI F1 `0.4897494305`.
+- The strict pure-empty gate covered Postel, Lommel and Arendonk across all tested thresholds and produced zero detections. Sparse-building contextual AOIs remained review-only evidence because real GRB buildings are present there.
+- Fixed-threshold persisted false-negative comparison used identical reference feature populations and reduced the false-negative rate in all seven positive AOIs versus the previous active `0.35` profile.
+- The guarded activation helper first returned `ready_to_apply`, resolved the exact local model asset and reported no downloads or inference. The reviewed `--apply` pass updated only `GEOINTEL_INSTALL_AI`, `YOLO_ENABLED`, `YOLO_MODELS_DIR` and `YOLO_MODEL_PATH` in the Tower environment.
+
+## UI and operator behavior
+
+- Detection Lab now recommends the promoted expanded-AOI model at threshold `0.15` and surfaces its seven-AOI coverage plus QA metrics.
+- The previous AOI1024 background-aware model remains mounted as an explicit legacy high-precision `0.35` review profile.
+- Profile selection remains deliberate: it selects a mounted local asset and run threshold only; it does not mutate runtime environment, download a model or start inference automatically.
+
+## Remaining limitation
+
+- Persistent small-building misses remain the weakest quality bucket. Continue with targeted evidence review and data coverage before considering another training run; do not infer production accuracy from aggregate F1 alone.