Record background-aware AOI1024 YOLO gate
This commit is contained in:
@@ -7,6 +7,17 @@
|
||||
|
||||
# Changelog
|
||||
|
||||
## Sprint 154 Background-aware AOI1024 YOLOv8s candidate gate (2026-07-09)
|
||||
|
||||
- Exported and audited background-aware AOI1024 training dataset `/app/storage/operator-data/yolo-building-aoi1024-bgaware512r3`; the audit passed with 162 tiles, 117 positive tiles, 45 negative tiles, 21,530 labels and no warnings.
|
||||
- Trained inactive local model asset `geointel-building-yolov8s-aoi1024bg512r3e50-pt` from the background-aware dataset. The Tower catalog reports SHA256 `e0980572aac90e7efc514608eb16d7de5bfbf27a4bbec04e7bc1bc8c02f9601f`, `status=available`, `active=false` and `will_download_models=false`.
|
||||
- Ultralytics validation for the completed 50-epoch CPU run ended at approximately precision `0.440`, recall `0.362`, mAP50 `0.251` and mAP50-95 `0.0878`.
|
||||
- Ran the seven-reference AOI1024 persisted QA matrix at tile `512`, overlap `64` and thresholds `0.25`, `0.15` and `0.05`. Mean positive F1 improved to `0.4908049127242224` at threshold `0.05`, `0.5074022485589402` at `0.15` and `0.44378879337957716` at `0.25`.
|
||||
- Ran an additional conservative-threshold positive matrix at thresholds `0.35`, `0.45` and `0.60`. Threshold `0.35` produced mean F1 `0.32086574003576274`, mean precision `0.840006` and mean recall `0.202135`.
|
||||
- Ran full hard-negative/background matrices. The candidate still fails automatic default promotion because mixed background-candidate AOIs keep false-positive pressure: max detections were `184` at threshold `0.05`, `103` at `0.15`, `75` at `0.25`, `55` at `0.35`, `38` at `0.45` and `18` at `0.60`.
|
||||
- Generated promotion reports at `artifacts/detection-model-promotion/aoi1024bg512r3e50-full/detection_model_promotion_report.md` and `artifacts/detection-model-promotion/aoi1024bg512r3e50-high-threshold/detection_model_promotion_report.md`; recommendation remains `none`.
|
||||
- No API contract, migration, provider fetching, fake detection data, model download behavior or active model default changed.
|
||||
|
||||
## Sprint 153 AOI1024 clean-label YOLOv8s candidate gate (2026-07-09)
|
||||
|
||||
- Audited AOI1024 label-quality variants after paged GRB reference regeneration and selected `/app/storage/operator-data/yolo-building-aoi1024-visible050-minpx8` for training because it passed the dataset audit while keeping the 512px runtime scale.
|
||||
|
||||
Reference in New Issue
Block a user