529 Commits

Include renames

Author	SHA1	Message	Date
JensandClaude Opus 5	08188005bd	correct the tiled inference chain and move runs off the request thread ... Tile handling produced results that were wrong before any model quality question arose: - orthophoto tiles reached the model through PIL convert("RGB"), which truncates the high byte of a 16-bit product and treats a 4-band RGB+NIR tile's infrared channel as colour. Tiles are now read with rasterio, the visible bands are chosen explicitly, and values are percentile-stretched across all three bands together so hue is preserved; - an object wider than the tile overlap was truncated by both tiles into two boxes that barely intersect, so IoU suppression kept both: two false positives and one missed footprint per seam building. Suppression now also compares overlap against the smaller box, and boxes cut by an interior tile edge are dropped in favour of the neighbouring tile's complete view; - georeferencing fell back to an assumed EPSG:4326 when a manifest carried no CRS, producing geometry that renders plausibly in the wrong place. QA already refused such a tile; inference now fails closed too. Segmentation QA scored candidates against every reference feature in the dataset, so every building outside the inferred tiles counted as a false negative. It now applies the same persisted tile coverage that detection QA has always used, including the indexed ST_Intersects prefilter. Duplicate suppression uses an STRtree instead of the O(n^2) scan, tiles are predicted in batches of YOLO_BATCH_SIZE (a setting that existed but was never read), and detection/segmentation runs can be queued through /run-async for a polling background worker rather than holding an HTTP worker thread for minutes of GPU work. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-22 14:32:44 +02:00
JensandClaude Opus 5	2b968b74cf	make QA scoring reproducible and threshold-independent ... The greedy IoU matcher gave a reference to whichever candidate was offered first. Row order decided that, and every detection in a run shares one transaction timestamp, so ordering by created_at left the assignment undefined: the same QA run over the same data produced different mean IoU, and the geometry shown to a reviewer as a false positive could be the better of two detections. Candidates are now ranked by confidence with feature identity as tiebreaker, which is also the COCO/PASCAL rule. A single precision/recall/F1 triple describes one operating point, so two models cannot be compared from it: a conservatively calibrated model looks worse at a low confidence cut and better at a high one without detecting anything differently. DetectionMetricsService adds the full curve, average precision and the threshold where F1 actually peaks. Also: - report the population the metrics were computed over, so matches + false_positives equals candidate_feature_count even under an area filter; raw dataset totals move to the _raw fields; - state whether candidates are axis-aligned boxes or footprint polygons. A box can never reach IoU 1 against a rotated building, so the strict score has a ceiling that has nothing to do with detection quality. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-22 14:31:21 +02:00
Jens	48084799c9	correct overlapping checkpoint evaluation	2026-08-09 20:41:54 +02:00
Jens	2de438b9cc	derive cleaner min-4px YOLO corpus	2026-08-09 19:51:17 +02:00
Jens	0209167cfd	fix(results): gate operational AI exports on QA	2026-08-09 11:31:51 +02:00
Jens	b76cd1837b	fix(ai): bind model scope to immutable geometry	2026-08-09 10:54:52 +02:00
Jens	70fb4b94e5	fix(accuracy): stabilize phase 4 evidence fingerprint	2026-08-02 04:53:16 +02:00
Jens	ee1982eff2	fix(accuracy): close phase 4 evidence bypasses	2026-08-02 04:29:10 +02:00
Jens	bd2fd9780f	feat(accuracy): build hardened phase 4 evaluation harness	2026-08-02 02:43:41 +02:00
Jens	7b92e29e49	feat(accuracy): complete phase 3 data scan and quarantine	2026-08-02 01:06:35 +02:00
Jens	5b3c17b494	feat(provenance): govern source snapshots and data inputs	2026-08-01 23:46:17 +02:00
Jens	9e582f2921	Align AI image packaging with PyTorch	2026-08-01 16:41:53 +02:00
Jens	24dec16744	Pin Tower-compatible CUDA runtime	2026-08-01 16:06:15 +02:00
Jens	dfcc11b0d5	Expand scoped demo analysis access	2026-08-01 16:00:13 +02:00
Jens	3effc77ed7	Add direct polygon label QA	2026-07-30 05:47:26 +02:00
Jens	08422ca50c	Add fail-closed source class and SAM fallback filters	2026-07-30 04:24:35 +02:00
Jens	29378bba42	Harden SAM roof refinement geometry gates	2026-07-30 04:00:24 +02:00
Jens	5ac3238c27	Reject same-year post-mosaic building labels	2026-07-30 03:35:15 +02:00
Jens	c2859e5919	Spread capped failure samples across weak contexts	2026-07-30 00:45:04 +02:00
Jens	15ff6f6433	Train against merged national roof instances	2026-07-30 00:10:15 +02:00
Jens	a5b321fdb6	Keep rejected training regressions from propagating	2026-07-29 23:33:53 +02:00
Jens	b8e40ba7a8	Prioritize recall in failure-driven sampling	2026-07-29 23:10:19 +02:00
Jens	16a1552427	Stabilize near-gate regional sampling	2026-07-29 21:24:50 +02:00
Jens	bb3df850aa	Rotate capped training samples across iterations	2026-07-29 20:59:03 +02:00
Jens	cfce3e64ec	Target related hard-negative contexts	2026-07-29 18:27:54 +02:00
Jens	a56df8b1ed	Cap regional failure oversampling	2026-07-29 17:51:32 +02:00
Jens	21c464c715	Require tile quality evidence for training	2026-07-29 17:37:55 +02:00
Jens	037f875575	Version rotated Belgian holdout corpora	2026-07-29 17:23:36 +02:00
Jens	8c9a58948b	Support governed train-only corpus shards	2026-07-29 17:21:18 +02:00
Jens	08d406ce7d	Expand failed regional building contexts	2026-07-29 17:14:17 +02:00
Jens	2fd9da9e16	Target regional training failures by context	2026-07-29 17:10:19 +02:00
Jens	e73468318f	Recover active training loop checkpoints	2026-07-29 16:32:01 +02:00
Jens	18387448de	Freeze aerial warmup across training loop	2026-07-29 16:27:30 +02:00
Jens	b393d9f65d	Continue rejected v37 through frozen loop	2026-07-29 16:23:28 +02:00
Jens	f18a23c060	Validate training completion handoff	2026-07-29 15:48:51 +02:00
Jens	d6502140da	Handoff completed training into closed loop	2026-07-29 15:45:41 +02:00
Jens	d5bf9c102e	Supervise checkpointed Tower training	2026-07-29 15:43:26 +02:00
Jens	1ce291d652	Gate completed checkpoints through training loop	2026-07-29 15:37:08 +02:00
Jens	b5155c702f	Automate failure-driven training continuation	2026-07-29 15:34:09 +02:00
Jens	e8530ae476	Unblock objective Belgian training loop	2026-07-29 15:31:55 +02:00
Jens	42da5b2c7f	Make Tower deploy remote explicit	2026-07-28 07:23:37 +02:00
Jens	438f10fb4b	Polish GeoIntel workbench and harden Tower deploy	2026-07-28 07:21:48 +02:00
Jens	c76a746cd7	Update	2026-07-27 23:28:43 +02:00
Jens	52addf2297	Update	2026-07-27 18:38:52 +02:00
Jens	100d9e220b	Balance positives during precision correction	2026-07-27 14:22:30 +02:00
Jens	9e2dde2b19	Preserve validation manifests in failure sampling	2026-07-27 12:45:08 +02:00
Jens	71c13c95c7	Rotate release holdouts after failed fixed test	2026-07-27 10:48:39 +02:00
Jens	51db5baa9c	Audit and suppress nested detector proposals	2026-07-27 09:58:25 +02:00
Jens	efd3272bda	Drive corpus sampling from calibration failures	2026-07-27 09:40:01 +02:00
Jens	1c6cf6a6a8	Make building training loop calibration-gated	2026-07-27 09:29:07 +02:00