Filter low variance YOLO negative tiles

scripts/README.md

@@ -367,6 +367,12 @@ negative tiles, and records `yolo_tile_dataset_summary.json` with
 `--min-label-visible-ratio` drops labels where only a small clipped fragment of
 the original building bbox is visible inside the tile; this reduces noisy
 tile-edge labels in overlapping-tile datasets. Use `0` for legacy behavior.
+Use `--drop-low-variance-negatives` to skip negative tiles whose rendered image
+has a max-min pixel range at or below `--blank-range-threshold`. This gate is
+intended for blank/no-data pure-empty negatives only; positive/labeled tiles are
+not removed by this filter. The summary records
+`skipped_low_variance_negative_tile_count` and skipped tile records with
+`skip_reason=low_visual_variance_negative`.
 For legacy operator manifests that predate explicit `background_category`, the
 exporter derives the same categories as the split-background evaluator:
 background samples with `reference_feature_count == 0` become
@@ -387,10 +393,16 @@ docker exec -it geointel python3 /app/scripts/export_operator_yolo_tile_dataset.
   --negative-keep-ratio 1.0 \
   --min-label-px 12 \
   --min-label-visible-ratio 0.35 \
+  --drop-low-variance-negatives \
+  --blank-range-threshold 3 \
   --val-samples turnhout,retie,westerlo,arendonk_heide \
   --force
 ```
 
+This refreshed cleanpx dataset is the minimum pre-training baseline after the
+visual contact-sheet pass found six blank-looking `arendonk_heide` validation
+negatives in the older export.
+
 Then audit with stricter small-box gates:
 
 ```bash