Jens/geointel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Expand operator samples for YOLO hard negatives
GeoIntel CI / docs-smoke (push) Has been cancelled
Details
GeoIntel CI / contract-smoke (push) Has been cancelled
Details

Browse Source
This commit is contained in:
Codex
2026-07-07 21:42:23 +02:00
parent e8d79fccbb
commit 89c5729d33
8 changed files with 288 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+13
View File
@@ -7,6 +7,19 @@
# Changelog
## Sprint 131 Operator sample expansion and negative-tile YOLO candidate (2026-07-07)
- Expanded `scripts/prepare_operator_real_data_samples.py` from the original Geel/Mol/Turnhout corpus to 7 reference AOIs plus 3 background-candidate AOIs.
- Added `sample_role` and `allow_empty_reference` metadata so deliberate background candidates can be prepared without weakening the empty-GRB guard for normal reference samples.
- Added regression coverage in `backend/tests/test_sprint131_operator_sample_expansion.py` for the expanded sample registry, empty-reference background candidates and normal reference-sample rejection.
- Live Tower preparation produced 10 operator samples: Geel, Mol, Turnhout, Herentals, Balen, Retie, Westerlo, Postel-bos, Lommel-heide and Kasterlee-bos.
- Live Tower tile export produced `/app/storage/operator-data/yolo-building-tile-expanded160` with 360 tiles, 260 positive tiles, 100 negative tiles and 11213 clipped building labels.
- Live Tower 50-epoch CPU training produced `/app/models/geointel-building-yolov8n-expanded160e50.pt`; the model catalog exposes it as `geointel-building-yolov8n-expanded160e50-pt` with SHA256 `bf6a5e8d25a62d784ee53764ea11d7ce89c4e7aeeac7588010e497b8d7dafb2b`.
- Live YOLO preflight loaded `geointel-building-yolov8n-expanded160e50-pt` successfully with `status=ready`, `model_load_ok=true`, `manifest_valid=true`, `tile_paths_exist=true`, `will_download_models=false` and `will_run_inference=false`.
- Live 45-run Geel/Mol/Turnhout/Retie/Kasterlee-bos QA matrix showed the expanded model is the best current candidate on dense building AOIs: best overall score was Geel at tile `640`, threshold `0.05`, precision `0.30333333333333334`, recall `0.14748784440842788`, F1 `0.1984732824427481`.
- Hard-negative finding: on the sparse Kasterlee-bos sample, `yolov8s-building-segmentation-pt` remained cleaner, while the expanded local model produced too many false positives. The model is therefore improved but still experimental, not a V1 default.
- No Training Studio UI, API contract change, provider fetching, model auto-provisioning, fake detections or app-side model training behavior was introduced.
## Sprint 130 Operator YOLO tile-level dataset tooling (2026-07-07)
- Added `scripts/export_operator_yolo_tile_dataset.py` to convert prepared operator samples into overlapping YOLO tile datasets with clipped building labels and deterministic negative tile retention.