Benchmark route · under review

CSTD: an exactly-two-region control route

A synthetic route where every image carries exactly two texture regions by construction, added to separate two failure modes that the real-world benchmarks leave tangled together: collapsing to one mask, and fragmenting into far too many.

Images 256 (hand-verified subset) Methods 5 Protocol no-oracle, region count withheld

Headline

2.00 Regions predicted by Detecture, against a ground truth of exactly 2. It is the only method of the five that lands on the right count.

256/256 Images on which Detecture returns exactly two regions. Every error on this route is a misplaced boundary, never a miscount.

0.0000 Boundary F-measure for LISA and GLaMM. Both collapse to a single mask, which leaves no internal boundary to score.

21% Acceptance rate when the released ground truth was checked by eye: 274 of 1,296 candidates reviewed.

The sharpest thing CSTD shows is that Hungarian mIoU alone is not a partition metric. LISA scores 0.8394 mIoU here, second only to Detecture, while covering 0.5094 of the image and finding no boundary at all: one mask that happens to overlap one of two regions is enough to inflate the score. Region Covering and boundary‑F separate them immediately.

Median mIoU across the 256 images is 0.9664. Full per-image results, including the failures, are in the gallery.