Refresh Belgian training gap evidence
This commit is contained in:
@@ -47,10 +47,14 @@ simultaneous gate.
|
|||||||
|
|
||||||
## Current gap inventory
|
## Current gap inventory
|
||||||
|
|
||||||
The v4 dated corpus closes basic composition, leakage and temporal-identity
|
The active v31 rotated-holdout corpus contains 142 independent AOIs and 26,041
|
||||||
gaps with 75 independent AOIs and 13,765 accepted building labels. It adds
|
accepted source building labels. Its automated corpus audit has no failures:
|
||||||
coastal, ribbon-development, farmland, park, forest and additional urban
|
spatial leakage is zero, temporal identity is explicit and 272 sub-resolution
|
||||||
contexts. Its remaining known gaps are:
|
plus 393 post-imagery labels are rejected rather than silently trained. The
|
||||||
|
tile-quality audit records 51,244 visible training instances, zero invalid or
|
||||||
|
missing label files, zero low-variance positive tiles and 214 negative tiles.
|
||||||
|
It covers coastal, ribbon-development, farmland, park, forest, industrial,
|
||||||
|
rail, quarry and additional urban contexts. Its remaining known gaps are:
|
||||||
|
|
||||||
- dated 2025 imagery is used for Flanders and Brussels and the dated 2024 SPW
|
- dated 2025 imagery is used for Flanders and Brussels and the dated 2024 SPW
|
||||||
campaign for Wallonia. Exact flight days remain a later metadata refinement,
|
campaign for Wallonia. Exact flight days remain a later metadata refinement,
|
||||||
@@ -58,8 +62,13 @@ contexts. Its remaining known gaps are:
|
|||||||
- GRB/PICC/UrbIS describe ground footprints, whereas visible roofs can remain
|
- GRB/PICC/UrbIS describe ground footprints, whereas visible roofs can remain
|
||||||
displaced. The existing detector QA contract therefore uses IoU 0.25; the
|
displaced. The existing detector QA contract therefore uses IoU 0.25; the
|
||||||
threshold is frozen and cannot be relaxed per candidate;
|
threshold is frozen and cannot be relaxed per candidate;
|
||||||
- the first loop candidate generalises poorly in Flanders and Wallonia,
|
- the active v38 loop still has to prove the frozen regional calibration gates;
|
||||||
especially Mechelen, Sint-Niklaas, Leuven, Mons and dense PICC areas;
|
iteration 2 improved aggregate calibration F1 to 0.573 but remained below
|
||||||
|
the Flemish F1/precision/recall gates and the Walloon precision gate;
|
||||||
|
- Oostende coastal fabric is the dominant observed Flemish calibration
|
||||||
|
failure. Independent train-only coastal positives and port/dunes negatives
|
||||||
|
are therefore selected by the failure-driven sampler without opening or
|
||||||
|
copying protected calibration data;
|
||||||
- sparse hard contexts pass the empty-image test more easily than dense urban
|
- sparse hard contexts pass the empty-image test more easily than dense urban
|
||||||
recall, so both gates must remain independent;
|
recall, so both gates must remain independent;
|
||||||
- building boxes are a valid first detector contract, but footprint-perfect
|
- building boxes are a valid first detector contract, but footprint-perfect
|
||||||
@@ -67,7 +76,7 @@ contexts. Its remaining known gaps are:
|
|||||||
|
|
||||||
The active production model remains unchanged while any gate fails.
|
The active production model remains unchanged while any gate fails.
|
||||||
|
|
||||||
The v31 expansion adds eighteen independent train-only AOIs after the v30
|
The v31 expansion added eighteen independent train-only AOIs after the v30
|
||||||
rotated-holdout calibration isolated the remaining domain gaps. Nine Flemish
|
rotated-holdout calibration isolated the remaining domain gaps. Nine Flemish
|
||||||
AOIs cover industrial roofs, ribbon development and coastal urban fabric. Nine
|
AOIs cover industrial roofs, ribbon development and coastal urban fabric. Nine
|
||||||
Walloon AOIs add dense urban, rural-town, regional-architecture and difficult
|
Walloon AOIs add dense urban, rural-town, regional-architecture and difficult
|
||||||
|
|||||||
Reference in New Issue
Block a user