Balance YOLO label QA coverage
This commit is contained in:
+3
-1
@@ -13,7 +13,9 @@
|
||||
- Preserved Turnhout, Retie, Westerlo and Arendonk-heide as manifest-backed validation holdouts and made the tile exporter reject unknown samples or holdout leakage.
|
||||
- Added `recommended_split` provenance to generated sample/reference/tile metadata and recorded the validation split in dataset summaries.
|
||||
- Hardened persistent false-negative comparison so portfolios with different reference feature identities cannot be compared.
|
||||
- Documented the expanded low-minimum-label dataset flow; no model was activated and no product API or migration changed.
|
||||
- Generated and audited the 20-source expanded dataset on Tower: 171 retained tiles, 45,892 valid labels, 9 low-variance negatives removed and no structural audit warnings.
|
||||
- Balanced visual label QA by source sample before selecting repeated dense tiles; the live pass covered all 19 retained sources without invalid labels, missing images or blank selections.
|
||||
- No model was activated and no product API or migration changed.
|
||||
|
||||
## Sprint 170 Persistent false-negative evidence audit (2026-07-12)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user