Document YOLO CRS live validation
GeoIntel CI / docs-smoke (push) Has been cancelled
GeoIntel CI / contract-smoke (push) Has been cancelled

This commit is contained in:
Codex
2026-07-07 02:11:46 +02:00
parent bc87681ab7
commit 673efd7a9d
2 changed files with 18 additions and 3 deletions
+1
View File
@@ -14,6 +14,7 @@
- Added regression coverage for the mixed-case YOLO class route that caused the Geel real-data smoke to persist zero detections.
- Confirmed through direct Tower inference that the active local building model returns raw detections on the prepared Geel orthophoto tile; the remaining work is threshold/QA calibration rather than model availability.
- Fixed raster tile manifest CRS propagation so generated tile manifests include source CRS metadata required to convert YOLO pixel boxes to WGS84 Detection GeoJSON coordinates.
- Deployed the class-normalization and tile-CRS fixes to Tower, reran the real-data detection + QA workflow, confirmed 4 persisted detections and verified Detection GeoJSON now returns WGS84 coordinates around Geel.
## Sprint 122 Real operator data availability and raster metadata fix (2026-07-07)
+17 -3
View File
@@ -27,16 +27,30 @@ Validation:
- quality check: `66907e6a-9ed7-4b0c-976f-5ad1ba9b8b7a`
- detection export: `516d37a3-2305-48a2-a3dd-b56f70eb055e`
- persisted detections used canonical `class_name=building` and preserved `model_class_name=Building`.
- Tower deploy from commit `bc87681` passed with `GEOINTEL_INSTALL_AI=true`.
- Deploy-time live migration smoke passed; PostGIS reported `3.6 USE_GEOS=1 USE_PROJ=1 USE_STATS=1` and Alembic head was `202606120900`.
- Deploy-time browser runtime verification passed for `http://192.168.10.150:1202`.
- Real-data smoke after the tile-CRS deploy passed and persisted 4 detections:
- project: `6ced26f3-4486-451e-a05e-26a8f859e361`
- raster dataset: `d54c3420-03d4-45a2-9501-bf1cce165a89`
- reference dataset: `319dad51-15e9-4ae6-815b-8e39023ed962`
- manifest: `/app/storage/tiles/6ced26f3-4486-451e-a05e-26a8f859e361/d54c3420-03d4-45a2-9501-bf1cce165a89/721c9557-5910-4994-b0ac-ba3418d12246/manifest.json`
- analysis run: `e99bcd07-ba48-4ebe-9ba9-24a94a8c1dcb`
- quality check: `32c8252b-fe60-4062-9639-125c62ca677f`
- detection export: `9d3a7f60-07a2-415c-9f75-a3c1988d0904`
- Verified the new tile manifest now carries `crs=EPSG:31370`, `source_crs=EPSG:31370`, `dataset_crs=EPSG:31370` and per-tile `crs=EPSG:31370`.
- Verified Detection GeoJSON now returns WGS84 coordinates around Geel instead of raw Belgian Lambert coordinates.
- Live detection QA persisted honest metrics for the Geel sample: 1 match, 3 false positives, 616 false negatives and score `0.0032206119162640897`.
Open:
- Redeploy the tile-CRS manifest fix and rerun the real-data smoke to verify Detection GeoJSON coordinates are in EPSG:4326.
- None for availability, persistence and CRS propagation.
Limitations:
- This fixes class routing, persistence and future tile manifest CRS propagation. Existing tile manifests generated before this fix remain missing CRS and should be regenerated before AI runs.
- Thresholds, false positives and reference IoU quality still need calibration on larger local orthophoto samples.
- The active building model is operational but not calibrated for production-quality local Belgian/Kempen orthophoto extraction. The Geel sample proves end-to-end persistence and QA, while the 1/617 reference match result shows that confidence thresholds, tiling strategy, class mapping and IoU defaults still need model-quality calibration.
Next recommended pass:
- Redeploy the tile-CRS fix, rerun the Geel real-data smoke, confirm WGS84 Detection GeoJSON coordinates and then calibrate practical confidence/IoU defaults.
- Run a detection calibration pass on several local orthophoto/reference samples: sweep confidence thresholds, inspect false positives/false negatives, tune tile size/overlap where needed and record a practical V1 baseline.
## Sprint 122 Real operator data availability and raster metadata fix (2026-07-07)