Prepare unique hard-negative training dataset

2026-07-09 02:20:49 +02:00
parent 54e6ad888e
commit 20c942df50
6 changed files with 39 additions and 5 deletions
@@ -444,11 +444,13 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Use live audit output to decide whether the next model pass needs more positive AOIs, label cleanup or unique hard negatives.
 - [x] Add more unique background/hard-negative AOIs before repeating hard-negative-balanced YOLO training.
 - [ ] Keep `yolo-building-tile-expanded160` as the clean current training baseline; avoid promoting r4/r8 repeat-heavy datasets as defaults.
- [ ] Regenerate Tower operator samples, export a new unique-hard-negative tile dataset and rerun the dataset audit before training.
+- [x] Regenerate Tower operator samples, export a new unique-hard-negative tile dataset and rerun the dataset audit before training.

 # Sprint 147 - Unique hard-negative AOI expansion

 - [x] Expand the documented operator background candidates from 3 to 9 unique AOIs.
 - [x] Keep every new background AOI explicit, `allow_empty_reference=True`, and `sample_role='background_candidate'`.
 - [x] Add test coverage for minimum background candidate count, unique centers and regional spread.
- [ ] Prepare the new samples on Tower and build a fresh hard-negative tile dataset.
+- [x] Prepare the new samples on Tower and build a fresh hard-negative tile dataset.
+- [ ] Rebuild Tower all-in-one image so the newly copied operator scripts are available inside `/app/scripts` without `docker cp`.
+- [ ] Train a new candidate from `yolo-building-tile-uniquehardneg160` and run the positive/background promotion gates before activating it.