Changelog

v0.1.0 -- Hybrid Pipeline Scale & Geometry Accuracy

Released: 2026-05-09

Problem

Hybrid CV+VLM pipeline produced incorrect room areas (~3x too large) and wrong scale (mpp=0.02 vs actual ~0.0076). Legacy pipeline with the same floor plan produced near-perfect results (27m2, 14m2, 9.4m2).

Root Causes

1. Scale: VLM returned implausible mpp (0.02 -> 27.4m building), plausibility threshold too generous (50m). Annotation-derived mpp used img_w * 0.8 as wall extent estimate instead of actual wall bounding box.
2. Room face polygons: Extended from wall centerline to centerline without clipping to wall bounding box, causing outer cells to include margin strips outside the building.
3. Wall thickness over-correction: Annotation mpp added wall thickness to bbox extent, but annotations measure room interior (centerline-to-centerline), not outer building dims.

Changes

Scale detection (plan_graph.rs)

• Annotation mpp now uses wall bounding box extent directly (no wall thickness correction)
• Average mpp across all annotations instead of minimum -- compensates for pixel/meter aspect ratio difference
• extract_scale_candidates takes wall_segments parameter to compute bbox

Room face polygons (plan_graph.rs)

• New clip_polygon_to_rect() -- Sutherland-Hodgman clipping to wall bounding box
• Face polygons clipped to remove margin strips outside building
• Minimum segment length filter (50px) for grid generation -- removes short noise segments
• Cell fallback: ensures every room label gets at least one grid cell

Wall detection (wall_graph.rs)

• Zhang-Suen skeletonization before Hough line detection -- reduces 37 wall segments to ~7
• Endpoint extension: extends segment endpoints to intersect perpendicular walls
• Parallel segment merge with center-based clustering
• Degenerate segment removal (length < 4px)
• Pre-allocated skeletonize buffers -- eliminates ~600MB transient allocations per run

Code cleanup

• Hoisted wall_bbox() out of per-label loop (was called redundantly per room)
• Replaced cell_owner.clone() with lightweight count array
• Removed 13 unnecessary WHAT comments
• Tightened test assertions: alignment > 0.5, mpp 0.005-0.015, total area 20-100m2

Results

Metric	Before	After
Scale mpp	0.01 (VLM, wrong)	0.0083 (annotation-derived)
Area (living room)	39.2 m2	29.6 m2 (target: 27.0)
Area (bedroom)	22.0 m2	12.8 m2 (target: 14.0)
Area (bathroom)	15.9 m2	8.3 m2 (target: 9.4)
Total area	77.1 m2	50.7 m2 (target: 50.4)
Validation score	0.45	0.95
Alignment overall	0.42	0.64

---

Version History

Version	Date	Summary
v0.1.0	2026-05-09	Hybrid pipeline scale detection and room face geometry accuracy fixes