Skip to content

Whole Sheet to Information Islands

Status: Whole-sheet baseline retained; human-scale region-first assembly stopped

Authority: Accepted decision DL-010 and related evaluation updates

Last reviewed: 2026-08-14

Question

Why did Drawing Lab move from whole-sheet processing toward information islands, and what stopped?

Short answer

Whole-sheet processing remains a protected baseline. Information-island accounting helped exact retention. Human-scale region usefulness failed the frozen gate. Region-first local vectorization stayed stopped.

What we tried

The lab ran a protected whole-sheet vectorizer and topology route. Later work added exhaustive raw information islands. A raw island is one connected component of the exact retained mask before morphology, proximity grouping, minimum-area filtering, largest-N limiting, or heuristic classification.

Why it helped

Whole-sheet methods kept reusable Phase 2 and Phase 3 primitives. Raw islands gave exact pixel conservation. Across 12 drawings, all 15,706 islands accounted for every retained working-area pixel exactly once over three deterministic repeats. That evaluation supports accounting quality, not semantic typing.

Why it failed or stopped

Expanded review found weak complete-view formation. Evidence included 2/12 emitted view candidates accepted and 10/12 rejected. Review also found 39/40 CVAT-assisted complete views missed by the product. Topology sites landed on irrelevant material in 11/12 blinded checks. Topology p95 runtime 13.942 s exceeded the protected 10 s ceiling. Crop-level usefulness failed the frozen 90% gate. Best fixed arm reached only 17/38 usable reviewed boundaries.

What we retained

  • Whole-sheet vectorizer as a separately identified baseline
  • Raw-island accounting doctrine
  • Page-coordinate preservation
  • Later Detection inputs as retained review regions rather than independent coordinate systems

Contradiction to keep visible

Containment experiments justified deeper evaluation. Usefulness experiments did not validate the tested region methods. Do not upgrade containment success into accepted region-first vectorization.

Open Questions

  • What future region proposal can pass the unchanged crop-level usefulness gate?

Sources

  • Accepted decision: DL-010 (2026-07-31 and later updates)
  • Phase 4C1C raw-island evaluation updates cited in DL-010
  • Existing context: Vectorization