CRAS real pair: independent registration checks and RGB point transfer (#4381)

This directory closes the real-pair gate of #4381 with measured numbers. It records what was done, exactly how, and what the held-out residuals do and do not support. Nothing here is an accuracy approval: the tolerance below was derived from held-out residuals after the solve, by a rule fixed beforehand, and every coverage number is reported as the planner measured it.

Source data and licence

• Scan and model: Abreu et al., CRAS laboratory record Zenodo 7948116, CC BY 4.0. Retain the attribution and the record DOI in any derivative. The paper's 3 mm figure concerns scan-to-scan registration and is not used here.
• Archive craslabannotated.zip, 4,267,281,425 bytes, downloaded complete on 2026-09-12 (curl -L -C - --retry 5, 2 min 37 s at ~26 MB/s); MD5 e5ecedab8f2a1d1f91861a3aec028a72 matches the record; SHA-256 933732e81910116cec5e74c9205074b82d1f571f3d9816386b68a0a909112671. Its single member CRASLAB_annotated.asc (26,752,658,348 bytes) holds a header row, a count row (584701977) and tab-separated X Y Z R G B Intensity Classification.
• Model craslabbim.ifc, 67,553,572 bytes, MD5 e20658f0d2d9e13c62363169b7fa3193 (matches), SHA-256 a83e5730…, IFC2X3 (see cras-ifc4-migration.json).
• No point data is committed. Archive, subsets and the full IFC live under ~/ci/cras/ on the host that ran this; every subset manifest pins SHA-256s.

Acquisition (whole archive, streamed, never unpacked)

1. python tools/texture-authoring/cras-archive-histogram.py craslabannotated.zip hist — one pass over all 584,701,977 rows (192 s): bounds x 0.01–18.32, y −4.18–20.27, z −2.07–3.25 m; 29,239 occupied 0.25 m cells; label counts. cras-archive-histogram.json. The floor sits at z ≈ −1.15 and the room ceilings at z ≈ 1.05 (corridor) / 1.55 (room B).
2. python tools/texture-authoring/cras-archive-subset.py craslabannotated.zip BOXES.json OUT — five further passes wrote spatially bounded, systematically decimated subsets as binary PLY (double XYZ in unchanged source metres, RGB8, label, original row index): subsets/ manifests give box, stride, points inside/kept and SHA-256 per file. Rooms were chosen from the occupancy map, not from file order: the corridor (x 7.5–10.75, y −4.5–0.5) and room B (x 10.75–18.5, y 2.75–11.25), plus dense strips along its north, east, south and west walls and two undecimated window-jamb boxes.

The IFC counterparts, from the published geometry (IfcOpenShell 0.8.3, world coordinates): the corridor between walls 0FDABnXCHCUR8gI54b$LZ2 (face x = −2.0197) and 1gsZ4XbYD0qRhfvwOUEK3Z (face x = 0.3907) with end wall 1gsZ4XbYD0qRhfvwOUEJ7O and its recessed door; room B between walls 1gsZ4XbYD0qRhfvwOUEI1z (south, y = 3.0085), 3qeiF73TD1uOeHIcNGcTy8 (north, y = 10.8615) and 0FDABnXCHCUR8gI54b$Kx2 (east brick wall, x = 7.9862) with five clerestory windows per drywall (sill 2.29 m, head 2.77 m) and five 0.5 m windows in the brick wall (sill 2.29, head 2.76). The IFC4 derivative (ifcopenshell.util.schema.Migrator, all 2,414 GlobalIds preserved) has identical wall geometry for every reference wall.

Landmarks: three-plane intersections, frozen fit/check lists

python tools/texture-authoring/cras-landmarks.py craslabbim.ifc subset landmarks.json (landmarks.json). Each landmark is the intersection of three named planar faces:

• IFC side: face coordinates read from the published IFC2X3 geometry (wall faces, opening jambs and heads, slab top z = 0.10). No IFC value is typed by hand; each row names the GlobalIds involved.
• Scan side: for each plane, points inside a search box are fitted by seeded RANSAC (600 hypotheses, 6 mm inliers, normal within 30° of the expected axis, ≥ 20 support points) followed by five least-squares refits; the box for the next plane is placed from the planes already fitted. A coarse offset (9.90, 0.22, −1.245) m read from wall-face histogram peaks places the first box only; it never enters a coordinate. Re-running every fit with the measured median offset instead moves no landmark by more than 1.7 mm (refit_with_measured_offset_drift_m), so the boxes did not choose the answer. Plane RMS is 0.6–3.5 mm; the SHA-256 of each plane's supporting archive rows is recorded.
• Partition: features are listed in a fixed order; those that could be measured alternate fit, check, fit, … starting with fit. The rule was written before any solve and the list was not revised against residuals.

Features and their fate:

Landmark	Planes	Result
Room B north clerestory band, west end, head	wall face ∩ jamb reveal ∩ head reveal	not measurable: the jamb there is a curved frame profile, no planar reveal
Room B north band, east end, head	north face ∩ east face ∩ head	fit
Room B south band, west end / east end, head	as above	check / fit
East brick wall, windows 1328061 · 1327917 · 1327933 · 1327949 · 1327965, S and N jambs at the head	east face ∩ jamb reveal ∩ head reveal	7 measured (fit/check alternating); 3 jambs are seen only at grazing angles and keep 2–33 candidate points without a coplanar set of 20, one of them even undecimated
Room B west door 1328449, south jamb at floor / at head	west face ∩ jamb ∩ floor / head	fit / check
Corridor floor corners SW / SE	two wall faces ∩ floor	check / fit
Corridor door 1328314 jambs at the floor, W / E	end-wall face ∩ jamb ∩ floor	check / fit
Room B floor SE	two wall faces ∩ floor	not measurable: benches hide the floor corner

Sills were deliberately not used: from inside the room every sill reveal is hidden behind the sloped bottom frame member (see the window section below).

Result: 16 landmarks, 8 fit + 8 held-out, spanning both rooms, three walls, two heights (floor 0.10 m and heads 2.76–2.77 m) and 12 × 15 m in plan.

Registration (canonical solver) and the tolerance rule

node tools/texture-authoring/cras-register-transfer.mjs landmarks.json craslabbim-ifc4-tessellated.ifc SUBSET.ply WALL_GUID OUT ARCHIVE_SHA256 runs IfcAPI.registerScanCorrespondences on the frozen lists (registration-request.json, registration-report.json). Source frame cras-archive-native-metres-v1 bound to the archive SHA-256; target frame ifc-world-z-up-metres bound to the target IFC SHA-256.

	RMS	median	90th pct.	max
Fit (8)	4.09 cm	2.92 cm	—	7.20 cm (B-south-band-E-head)
Held-out (8)	4.53 cm	3.35 cm	6.54 cm	6.54 cm (B-west-door-S-head)

Rotation 0.72°; non-collinearity spread ratios 0.163 (source) / 0.165 (target).

Tolerance rule, fixed before the run: the largest held-out residual rounded up to the next centimetre → 7 cm. Fit residuals never enter it. This is the accuracy the transfer may claim for this pair: a scan surface within 7 cm of its IFC face is consistent with the independently checked registration; nothing tighter is supported. The residuals are not scanner noise (plane fits are millimetre-level) but as-built deviations from the model: the corridor is ~9 cm narrower than modelled (after registration its east face lies 6.0 cm into the corridor from the IFC face, its west face 3.1 cm), the north clerestory band starts ~10 cm further east than the modelled opening, and room B is 3–5 cm deeper between its south and north faces than modelled. A rigid fit spreads these over its 16 points; a transfer bound tighter than a wall's own deviation refuses that wall entirely (see the half-tolerance control).

Where the registered scan lies against each transfer wall

node tools/texture-authoring/cras-face-depth.mjs SUBSET.ply registration-report.json craslabbim-ifc4-tessellated.ifc WALL_GUID x|y OUT.json applies the registration to a subset and histograms, at 1 cm, every point in a wall's footprint along its thickness axis, measured from the wall's minimum face (face-depth-wall-1117.json, face-depth-wall-1273.json). This is the measured basis for the side statements below; it never enters a transfer.

Wall (10 cm)	scan surface	position after registration
Corridor drywall 1gsZ4XbYD0qRhfvwOUEK3Z, faces x 0.3907 / 0.4907	corridor face (the only one in the subset), 89.8 % of 52,588 points	6 cm in front of the modelled corridor face (x − 0.06), outside the solid
North wall 3qeiF73TD1uOeHIcNGcTy8, faces y 10.8615 / 10.9615	room B face, 52 % of 421,312 points	5–6 cm in front of the modelled south face (y − 0.05 … − 0.06), outside the solid
	far (north) side face, 22 %	4–5 cm inside the modelled solid (y + 0.04 … + 0.05): 5.5 cm behind the modelled north face, past its 5 cm midplane, and 4.5 cm behind the south face

So the real north wall is shifted ~5.5 cm south of its model: the room-B capture sits in front of the south face while the far side's capture sits inside the modelled wall, nearer to the south face than to the north face it belongs to. Without orientation this is the hardest side question the planner faces, and the transfer below reports it as such.

Transfer target: a tessellated derivative

The published model is swept solids with 1,328,372 entities (each manufacturer window/door family carries ~47,000 vertices). The F4 target policy takes direct IfcTriangulatedFaceSet bodies (evaluated occurrences are F6, #4404) and the planner bounds its target at 200,000 entities, so python tools/texture-authoring/cras-tessellated-target.py craslabbim.ifc craslabbim-ifc4-tessellated.ifc cras-ifc4-tessellated.json writes what an IFC4 Reference View export would: all 61 walls and 5 slabs as world-coordinate tessellated bodies (openings applied by the independent IfcOpenShell tessellation), GlobalIds and names unchanged, one neutral surface style each, 484 entities, 73,495 bytes (craslabbim-ifc4-tessellated.ifc, cras-ifc4-tessellated.json). The tessellated wall corners equal the extrusion corners the landmarks were read from.

RGB point transfer: measured coverage

Both runs use orientation: target-referenced (the archive carries no normals or stations), support radius 4 cm, 6–48 supporting points, 6 mm surface band, minNormalDot 0.8, ambiguity 2 mm, and maxDistanceMetres = maxBehindMetres = the tolerance. Times are single Node runs on this host, not a browser claim.

The half-tolerance control runs at exactly half the tolerance (3.5 cm), as cras-register-transfer.mjs computes it; the committed summaries carry maxDistanceMetres = maxBehindMetres = 0.035 for those rows.

Corridor drywall 1gsZ4XbYD0qRhfvwOUEK3Z (10 cm, no openings; source: corridor subset, stride 24, 465,029 points) — transfer-summary-wall-1117.json:

bound	texels/m	samples	observed	too far	normal	ambiguous	behind	sparse	observed area	work used
7 cm	64	105,126	33,843	63,591	793	700	119	6,080	8.20 m² of 25.28 m²	43.5 M of 128 M (376 ms)
3.5 cm	64	105,126	0	105,092	0	15	0	19	0	15.8 M (113 ms)

The wall's 25.3 m² are two 4.1 × 2.9 m faces plus edges. Only the corridor face was scanned in this subset, and only below the corridor's suspended ceiling (scan z 1.04 ↔ IFC 2.29 m), so ≈ 9.1 m² could ever be observed; 8.20 m² were. Per face from the independent reopen (observed_by_face_normal in oracle-wall-1117.json, 55 barycentric samples per triangle): corridor face (−x) 75.5 % of samples observed, far face (+x) 0 %, bottom edge 15.5 %, every other face 0 %. The far face, the 0.7 m above the ceiling and the edges are unknown by distance. At half the tolerance nothing is observed, because the as-built face lies 6 cm in front of the modelled one — the tolerance is doing exactly the work the held-out residuals licence. Atlas 512 × 700 (textures/b12ae7a6….png); the bright band near the ceiling is the fluorescent lighting in the capture.

North wall 3qeiF73TD1uOeHIcNGcTy8 (10 cm drywall with five clerestory windows, both faces scanned; source: strip stride 36, 623,244 points) — transfer-summary-wall-1273.json:

bound	texels/m	outcome
7 cm	64	refused: Transfer work budget exhausted; reduce source extent or atlas density (763 ms)
7 cm	32	40,672 samples, 16,033 observed, 7,975 too far, 3,024 normal, 1,256 ambiguous, 7,478 behind, 4,906 sparse; 15.14 m² of 37.22 m²; 59.5 M work (331 ms); atlas 512 × 374
3.5 cm	32	1,521 observed, 33,898 too far, 945 behind; 1.34 m²

Per face from the independent reopen (oracle-wall-1273.json): room-B face (−y) 49.0 % of samples observed, far face (+y) 7.4 %, window heads / bottom (−z) 13.8 %, jambs (±x) 21–24 %, tops (+z) 0.5 %. This is the side question from the depth table resolved by target geometry alone: the room-B face takes the room capture 5–6 cm in front of it (a capture in front of a face is preferred over one inside the solid), and the far side's capture, 5.5 cm inside the modelled wall and past its midplane, is refused as behind on the north face instead of being painted there — the 7,478 behind samples. The 7.4 % the north face does observe is what lies within 7 cm in front of its modelled position (a surface at y + 0.15 … + 0.16 in the depth histogram). Before the nearest-face rule this wall reported 21,892 observed / 292 behind with both faces ~39 % observed, i.e. the north face was painted from a capture 5.5 cm behind it and the room-B face partly from the far side's capture 4.5 cm behind it; those numbers were withdrawn with this revision.

The refusal at 64 texels/m is the bounded job behaving as specified on a real wall: two scanned faces 10 cm apart double the points every query visits, and the fixed 128,000,000-unit budget stops the plan rather than degrading it. The planner now reports budget.workUsed beside coverage so this is auditable.

Independent apply and reopen

python tools/texture-authoring/cras-transfer-oracle.py craslabbim-ifc4-tessellated.ifc RUN tolerance WALL_GUID OUT applies only the plan's created/edited/removed entities with IfcOpenShell, writes the transferred IFC and PNG, reopens the file and checks it (oracle-wall-1117.json, oracle-wall-1273.json): wall geometry unchanged (12 / 60 triangles), exactly one IfcIndexedTriangleTextureMap indexing every triangle, the image URI equal to the atlas asset (named by its own SHA-256), IfcSurfaceStyleWithTextures assigned, IfcOpenShell schema validation 0 errors (EXPRESS rules not run). Sampling every triangle's UV centroid in the atlas: 2 of 12 (the corridor face) and 12 of 60 triangles carry captured colour, the rest the prior style; observed_by_face_normal samples 55 barycentric points per triangle and groups them by the triangle's world normal (the per-face percentages quoted above).

apps/viewer/src/lib/appearance/scan/cras-transferred-output.test.ts is the ifc-lite side over the committed craslabbim-ifc4-tessellated-transferred-wall-1117.ifc: the wall tessellates with identical ordered corners, one UV per vertex and the atlas reference; packaged as IFCZIP (fflate + unwrapIfcZipWithResources) the model bytes and the texture round-trip byte-exact; seeded into a room by the real owner seed and reconstructed by a fresh guest, the wall is selectable by its GlobalId path with byte-exact texture pixels and UVs.

What this does and does not establish

• Established: a licensed real scan/IFC pair, sixteen independently measured three-plane correspondences with frozen fit/check lists, held-out residuals published before any bake tolerance was chosen, a tolerance derived from them by a pre-stated rule, and RGB point transfer onto two walls of that model with observed/refused/unknown coverage, budget refusal, independent reopen, IFCZIP packaging and room join all measured on the output.
• Not established: absolute survey accuracy (the landmarks are manually defined building features measured in both datasets, not surveyed control points); any tolerance below 7 cm for this pair; behaviour on voided extrusions (the target is a tessellated derivative); browser timing (Node runs only); the colour fidelity of the captured atlas beyond what the scanner recorded; the side of an unoriented capture that lies deeper inside a modelled wall than its midplane (by geometry it belongs to the opposite face, and nothing in this archive can overrule that).
• Orientation: with no normals or stations in the archive, every point plan is target-referenced, and the side of each sample rests on the target's own geometry: self-occlusion for captures outside the solid, nearest-face attribution for captures inside it (behind bound capped at half the thickness), and preference for a capture in front of the face over one inside the solid. The controlled and WASM tests pin these with captures 1 mm outside the faces (all three orientation sources), 1 mm and 3 mm inside a 4 mm partition, and the measured 10 cm-wall configuration above (5.5 cm in front / 4.5 cm inside).