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

Balance YOLO label QA coverage
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
2026-07-12 23:53:18 +02:00
parent 0f49c980ba
commit 3266c6578d
6 changed files with 117 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/CODEX_EXECUTION_LOG.md
+18 -2
View File
@@ -6991,11 +6991,27 @@ Open:
- 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.
## Tower runtime evidence
- Pushed `0f49c98` and redeployed the all-in-one runtime at `http://192.168.10.150:1202`; live migration and browser proxy verification passed.
- Refreshed `/app/storage/operator-data/operator-samples-1024/operator_samples_manifest.json` to schema version 2 with 20 sources.
- Newly fetched real GRB reference counts: Olen 1,952, Lille 1,839, Oud-Turnhout 2,691 and Kasterlee 1,831; all are `recommended_split=train`.
- Exported `/app/storage/operator-data/yolo-building-aoi1024-expanded-minpx4vis035`:
- 171 retained tiles;
- 153 positive and 18 negative tiles;
- 45,892 labels;
- 144 train and 27 validation tiles;
- 9 low-variance negatives skipped;
- validation holdouts Turnhout, Retie, Westerlo and Arendonk-heide recorded in summary provenance.
- Dataset audit status `ok`: no warnings, invalid labels or missing files; median normalized box area `0.000694274766`, small-box share `0.3832694151486098`.
- Improved the visual contact-sheet selector after the first live sheet overrepresented dense AOIs. The balanced rerun selected 40 tiles across all 19 retained source samples with zero invalid labels, missing images, missing label files or low-variance selections.
- Started one 50-epoch CPU YOLOv8s candidate as inactive runtime evidence: `geointel-building-yolov8s-aoi1024expandedminpx4vis035e50.pt`.
## 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.
- The inactive training run must finish and pass positive-AOI plus split-background promotion gates before it can be considered for activation.
## 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.
- Finish the inactive candidate, run fixed-threshold positive evidence and split pure-empty/sparse-context background matrices, and preserve the current production default unless the promotion report passes every gate.