Evidence-gated multimodal parsing and vectorization of architectural floor plans

Hongxuan Chen; Wenda Wang; Jiachen Lu; Qirui Shen; Zilong Huang; Lei He; Xinyue Dong; Weixin Huang

cs.CVSep 22, 2026

Evidence-gated multimodal parsing and vectorization of architectural floor plans

Authors: Hongxuan Chen, Wenda Wang, Jiachen Lu, Qirui Shen, Zilong Huang, Lei He, Xinyue Dong, Weixin Huang

Organizations: School of Architecture, Tsinghua University

Abstract

Architectural floor plans remain a high-friction barrier to archive digitization and early design-model preparation because heterogeneous graphics encode spatial semantics and editable geometry together. We introduce SALI-FP, an evidence-gated multimodal pipeline that converts a plan into reviewable semantic maps, objects, vectors, and relation records while constraining local revisions by image evidence. In a full production audit of 11,534 heterogeneous plans, SALI-FP produced structured outputs for every plan, including 752,510 valid polygon-bearing objects. The same output form has supported initial drawing digitization and design-model preparation in practical design work. Public-benchmark calibration is paired with a 30-case matched visual evidence set in Appendix F, where room-scale coverage, openings, oblique boundaries, and circulation continuity can be inspected directly. SALI-FP offers an engineering-oriented interpretation-to-geometry workflow for reviewed CAD/BIM preparation and existing-building information recovery.

Figures & tables

Fig. 1: Floor-plan recognition and vectorization: (a) source raster; (b) semantic output; (c) sparse inverse rendering.

Route	Main output	Comparison boundary
Supervised segmentation [ 2 , 5 ]	Pixel classes	Shared labels, canvas, and test split
Junction or polygon recovery [ 3 , 4 , 10 ]	Structured geometry	Native decoder and topology must be identified
Graph-based interpretation [ 6 , 7 ]	Objects and relations	Align object and relation scope
Vision-language understanding [ 12 , 13 , 14 , 15 ]	Semantic or vector predictions	Evaluate localization explicitly
Tool-using agents [ 16 , 17 , 18 ]	Plans, actions, and state	Report execution and drawing evidence separately

Table 2

Fig. 2: SALI-FP architecture and evidence flow: (a) global semantic parsing; (b) evidence-gated revision and protected candidate selection; (c) representation generation and a connected output package. Rounded containers group operations; rectangular frames contain data and recorded evidence. Arrows identify data or selection conditions. The frozen CubiCasa5K test 0002 example supplies the images and E01 record; three accepted proposals illustrate the recorded evidence-gated execution. G retains generated relations. The local geometric extension is evaluated as a subsequent representation stage.

Condition	Setting and action
Confidence / target support	Reject below 0.80 / below 16 target pixels
Existing target label	Reject at 0.70 coverage; construction targets at 0.15
Other named room / blank or fallback	Reject at 0.30 other-room coverage / below 0.35 support
Source structure	Intensity below 170; require 0.02 dark support for non-door construction
Image-change fraction	Repair at most 0.55; cleanup at most 0.35
Cleanup foreground change	Non-white fraction changes by at most 0.15

Table 4

Fig. 3: Recorded stages of cubicasa5k_test_0001: (a) source; (b) initial candidate; (c) protected repair; (d) cleanup; (e) final source-aligned labels.

Fig. 4: Semantic-to-Rhino interface and retained model illustrations.

Item	Record
Working plans / format	11,534 / 1024 × 1024 PNG
Development / validation / test	9,225 / 1,157 / 1,152
Hard subset	400 within test
Independent-reference annotation package	200 stratified tasks
Release governance	Source-specific access, privacy, and redistribution review

Table 7

Method	Preprocessing and decoder	Evaluation scope
CubiCasa5K [ 2 ]	Four right-angle rotations; full 44 channels	Four-class argmax; official polygon decoder
Raster2Seq [ 27 ]	256-pixel bicubic input; released EMA checkpoint	Native room polygons; room, corner, angle matching
MiT-UNet [ 25 ]	512-pixel native normalization; wall threshold 0.5	Visible walls only
SALI-FP	Retained image/audit outputs; frozen representation stage	Four common classes; stored sparse room geometry

Table 8

Output layer	Delivered evidence	Scale	Engineering interpretation
Semantic labels $S$	Nonempty normalized semantic map	11,534 / 11,534 plans	Common input to object and geometry extraction
Objects $O$	Typed shape records	783,299 shapes	Explicit semantic units for review
Native vectors $V$	Polygon-bearing objects passing the delivery audit	752,510 / 783,299 (96.07%)	Editable geometry with retained status
Relations $G$	Generated relation pairs	1,591,739 pairs	Preserved structural context for downstream inspection
Trace $H,z$	State and recovery records	11,534 plans	Recoverable execution and review path

Table 9

Fig. 5: Evidence route for SALI-FP. Strict public-test calibration, full-corpus structured output, and fixed visual comparisons answer complementary parts of the evaluation. Detailed baseline scores and diagnostics are retained in Appendix C; the 30 matched complex-plan comparisons are in Appendix F.

Fig. 6: Entry points to the fixed visual-evidence set. Each source/SALI-FP pair corresponds to a full, matched baseline comparison with two aligned detail regions in Appendix F.1, F.7, F.16, and F.30; Appendix F contains all 30 cases.

Fig. 7: ArchP10k structured-output audit: (a) delivered output layers; (b) polygon-bearing geometry passing the delivery audit; (c) per-plan representation scale; and (d) retained audit-action records. Quantities have different denominators and are presented separately.

Appendix figures & tables48 assets

Supplementary material from the paper’s appendix.

Appendix

Class	RGB	Common target
wall	255, 0, 0	Structural boundary
door	255, 255, 0	Door/opening
window	0, 200, 83	Window
stair	0, 102, 255	Background (unmapped)
elevator	255, 128, 0	Background (unmapped)
living_room	0, 255, 255	Room

Appendix

Table 13

Source group	Plans	room_default / predicted non-background pixels
Residential	7,210	13.6%
Commercial and hospitality	942	15.1%
Office, education, and research	1,169	28.0%
Mixed use	305	20.4%
Civic, cultural, and health	941	27.2%
General interior	678	14.8%

Appendix

Table 14

Predicted class	Pixels	Share of predicted non-background pixels
living_room	1,419,369,397	24.21%
room_default	987,723,922	16.85%
bedroom	764,413,067	13.04%
wall	699,703,665	11.94%
balcony	487,712,685	8.32%
kitchen	349,764,932	5.97%

Appendix

Table 15

Item	Count
Original sparse objects	783299
Originally valid	673880
Valid polygon-containing objects after repair	752510
Accepted repairs	78630
Raster-guard rejection	22963
Collapsed without polygon	7820

Appendix

Table 16

Setting	Value
Working resolution / ECC iterations / termination	512 pixels / 150 / 0.00001
Affine singular-value range / maximum ratio	0.65-1.35 / 1.25
Maximum rotation / corner displacement	5 degrees / 25% of diagonal
Foreground retained / source-score gain	At least 98% / strictly positive
Calibration source-gain candidates	0 / 0.01 / 0.03 / 0.06
Calibration selection	Maximum pooled mIoU; BIoU decline at most 0.005

Appendix

Table 17

Setting	Value
Input	Same 400 test plans; final SALI-FP and native baselines
Extraction	Pixel-cell polygons; four-connectivity; retain holes and components
Simplification grid (pixels)	0, 0.25, 0.5, 1, 2, 4, 8; preserve topology
Fidelity thresholds	0.95, 0.98 (primary), 0.995; minimum nonempty-class IoU
Selection	Fewest vertices; then bytes; then tolerance
Serialization	Compact UTF-8 JSON; six decimal places; identical fields

Appendix

Table 18

Step	Evidence and decision
Schema and location	Require source-coordinate bbox, valid palette target, confidence and evidence fields
Confidence and area	Require confidence >= 0.80 and target mask >= 16 pixels
Current target protection	Reject existing target coverage >= 0.70; construction threshold 0.15
Named-room protection	Reject other named-room fraction >= 0.30
Room support	Require blank or fallback support >= 0.35
Construction support	Require source intensity < 170 over >= 0.02 of the region; door exception

Appendix

Table 19

Field	Meaning
bbox; source_coordinate_space	Rough source-image region; image1_source_plan
semantic_key; expected_semantic_key	Requested class from the frozen palette
current_semantic_key	Candidate class or white_or_unmapped / unknown
first_pass_problem; problem_evidence	Visible candidate error and supporting observation
source_evidence; fix_rationale	Drawing cue and rationale for the proposed correction
confidence; source_cues; reason	Model assessment and concise textual support

Appendix

Table 20

Stage	Rule
1. Palette thresholds	Per-channel tolerance 20; minimum squared RGB distance; current and legacy aliases
2. White mixtures	Unassigned chromatic pixels; t in [0.06, 1.05]; off-ray <= 24; span >= 4
3. Ordered hue predicates	Unassigned, non-background pixels; value >= 40; span >= 35; Table B.7
4. Background	All remaining pixels receive integer ID 0
Frozen provenance	All-RGB LUT and classifier source hashes; stored pointset configuration per plan

Appendix

Table 21

Order	Target	Predicate
1	wall	((H <= 15) or (H >= 350)) and (R > 180) and (G < 100) and (B < 100)
2	door	(H >= 45) and (H <= 68) and (R > 180) and (G > 180) and (B < 100)
3	window	(H >= 85) and (H <= 150) and (G > 100) and (R < 110) and (B < 130)
4	stair	(H >= 200) and (H <= 245) and (B > 120) and (R < 100) and (G < 160) and (maxRGB >= 170)
5	elevator	(H >= 20) and (H <= 44) and (R > 180) and (G > 70) and (G < 180) and (B < 120)
6	living room	(H >= 170) and (H <= 192) and (R < 90) and (G > 190) and (B > 190) and (maxRGB >= 220)

Appendix

Table 22

Operation	Saved setting
Native contour simplification / minimum area	1.2 px / 24 px squared
Rectangle eligibility	Area ratio >= 0.9; at most 6 approximated corners
Coordinate offset / short-edge length	0.5 / 2.0 px
Native orthogonal snap	1.5 px
Auxiliary room / construction simplification	0.6 / 4.0 px
Auxiliary expansion: room / construction / opening	1 / 1 / 1 px

Appendix

Table 23

Method / stage	Metric	Estimate [95% BCa CI]
SALI-FP / raw	oa	0.7830 [0.7757, 0.7899]
SALI-FP / raw	mean_pa	0.3001 [0.2903, 0.3123]
SALI-FP / raw	miou	0.2074 [0.2000, 0.2162]
SALI-FP / raw	biou	0.0570 [0.0530, 0.0623]
SALI-FP / raw	Visible-wall IoU	0.1116 [0.1013, 0.1245]
SALI-FP / raw	Wall clDice	0.1728 [0.1589, 0.1892]

Appendix

Table 24

Partition	Plans	Mean RCR [95% CI]	Micro RCR
full	11534	0.9569 [0.9565, 0.9573]	0.9595
development	9225	0.9571 [0.9566, 0.9575]	0.9597
validation	1157	0.9566 [0.9552, 0.9578]	0.9592
test	1152	0.9556 [0.9541, 0.9569]	0.9582
hard	400	0.9507 [0.9483, 0.9530]	0.9531
nonhard test	752	0.9582 [0.9565, 0.9596]	0.9612

Appendix

Table 25

Quantity	Value	Interpretation
Logical calls / HTTP attempts	39,875 / 41,931	2,056 excess request attempts
Audit median [IQR]	39.3 [24.0, 62.5] s	Recorded remote stage
Conditional repair median [IQR]	164.8 [100.9, 235.0] s	Only triggered repairs
Cleanup median [IQR]	138.4 [93.8, 211.5] s	Recorded remote stage
Recorded subtotal median	260.8 s	Not full end-to-end time
Reference-rate total / mean	$9,430.83 /$ 0.818	Rate-card calculation, not billing

Appendix

Table 26

Quartile	Plans	Mean RCR	Protected (%)	Median points
1	2884	0.9607	15.46	651.0
2	2883	0.9578	18.38	827.0
3	2883	0.9561	19.11	999.0
4	2884	0.9529	22.68	1288.0

Appendix

Table 27

Quantity	Count
Structured outputs	11534
One repair	5273
No repair	6261
Protective status	2181
Suggested / accepted / rejected	11,481 / 8,033 / 3,448
RCR below 0.90	191

Appendix

Table 28

Fig. C.1: Systematic-error examples: source, GT, final SALI-FP, and selected error pixels. Each case maximizes one absolute error count. Test 0048 lacks Space references; its room-addition mask also reflects annotation scope (Table C.23 ).

Quantity	Measured value
Exported plans	11534
Registration accepted	10981
Whole-plan geometry fallback	14
Objects receiving affine numerical repair	72
Plans with out-of-canvas coordinates	597
Mean / micro RCR, repaired geometry before registration	0.9290 / 0.9341

Appendix

Table 30

Method / stage	0.25%	0.50%	1.00%
SALI-FP / raw	0.1174	0.2100	0.3271
SALI-FP / registered	0.3264	0.4962	0.6131
CubiCasa5K / raw	0.8300	0.8888	0.9127
CubiCasa5K / registered	0.7820	0.8821	0.9108

Appendix

Table 31

Method / stage	Class	IoU	Dice
SALI-FP / raw	Boundary	0.1094	0.1972
SALI-FP / raw	Door/opening	0.0396	0.0762
SALI-FP / raw	Window	0.0520	0.0989
SALI-FP / raw	Room	0.6284	0.7718
SALI-FP / registered	Boundary	0.3180	0.4826
SALI-FP / registered	Door/opening	0.1454	0.2539

Appendix

Table 32

Fig. C.2: All-method sensitivity to identical source-only registration. Gray circles denote original outputs; blue diamonds denote registered outputs. Unmodified supervised predictions remain the primary baseline. MiT-UNet is included only for visible-wall metrics.

Method	GT wall IoU	Vertices [IQR]	KiB [IQR]	Achieved
SALI-FP	0.3325	644.0 [422.0, 967.0]	12.3 [7.7, 18.3]	400/400
CubiCasa5K	0.7482	2,912.0 [2,263.5, 3,574.5]	47.5 [37.5, 57.7]	400/400
MiT-UNet	0.7721	431.0 [309.5, 570.5]	6.5 [4.6, 8.8]	400/400

Appendix

Table 34

Target	Baseline	Cost	Ratio [95% BCa CI]	Paired
common4	CubiCasa5K	vertices	0.639 [0.610, 0.678]	400/400
common4	CubiCasa5K	bytes	0.886 [0.824, 0.941]	400/400
wall	CubiCasa5K	vertices	0.217 [0.202, 0.232]	400/400
wall	CubiCasa5K	bytes	0.245 [0.233, 0.269]	400/400
wall	MiT-UNet	vertices	1.438 [1.353, 1.534]	400/400
wall	MiT-UNet	bytes	1.692 [1.599, 1.797]	400/400

Appendix

Table 35

Group	Transition	Metric	Change [95% BCa CI]	Plans
all	protected_repair minus initial	plan-mean miou	-0.0088 [-0.0131, -0.0045]	400
all	protected_repair minus initial	plan-mean biou	-0.0032 [-0.0060, -0.0005]	400
all	cleanup minus protected_repair	plan-mean miou	0.0055 [0.0032, 0.0078]	400
all	cleanup minus protected_repair	plan-mean biou	-0.0040 [-0.0055, -0.0025]	400
all	final minus cleanup	plan-mean miou	0.1699 [0.1579, 0.1821]	400
all	final minus cleanup	plan-mean biou	0.1288 [0.1195, 0.1383]	400

Appendix

Table 36

Method	Metric	F1 [95% BCa CI]	Mean P / R	Mean plan F1
SALI-FP	Room	0.4450 [0.4221, 0.4683]	0.3647 / 0.5705	0.4271
SALI-FP	Corner	0.1590 [0.1494, 0.1690]	0.0973 / 0.4345	0.1559
SALI-FP	Angle	0.0932 [0.0863, 0.1002]	0.0573 / 0.2495	0.0916
CubiCasa5K / polygons	Room	0.6305 [0.6093, 0.6523]	0.6446 / 0.6170	0.6175
CubiCasa5K / polygons	Corner	0.3274 [0.3134, 0.3424]	0.2429 / 0.5020	0.3167
CubiCasa5K / polygons	Angle	0.2358 [0.2237, 0.2494]	0.1763 / 0.3563	0.2278

Appendix

Table 37

Method	Tolerance (px)	GT mIoU [95% BCa CI]	Median vertices	Completed
sali_fp	0	0.3596 [0.3449, 0.3740]	3,869.0	400/400
sali_fp	0.25	0.3596 [0.3449, 0.3740]	3,869.0	400/400
sali_fp	0.5	0.3596 [0.3449, 0.3740]	3,641.0	400/400
sali_fp	1	0.3613 [0.3464, 0.3757]	1,916.5	400/400
sali_fp	2	0.3612 [0.3463, 0.3756]	1,734.0	400/400
sali_fp	4	0.3609 [0.3461, 0.3753]	1,675.0	400/400

Appendix

Table 38

Fig. C.3: GT accuracy versus shared-decoder control-point cost at the seven fixed simplification tolerances in Table C.13 . Every setting is applied to all test plans; no per-plan selection uses GT. Markers denote the frozen tolerance settings.

Observation	Result	Scope
Plans with extra HTTP attempts	1,591 / 11,534	Per-case request counters
Additional HTTP attempts / maximum per plan	2,056 / 6	Not extra logical repairs
Repair-triggered / skipped public plans	263 / 137	Same 400 test plans
Already-target threshold 0.56	350 / 3760	One nonconstruction predicate only
Already-target threshold 0.7	200 / 3760	One nonconstruction predicate only
Already-target threshold 0.84	64 / 3760	One nonconstruction predicate only

Appendix

Table 40

Method	Completed	Cost n	Vertices [IQR]	KiB [IQR]	Filtered room F1
SALI-FP	400/400	398	294 [199, 431]	5.219 [3.596, 7.622]	0.4443
CubiCasa5K / polygons	395/400	393	132 [92, 188]	2.266 [1.619, 3.136]	0.6273
Raster2Seq	400/400	398	70 [52, 103.75]	1.457 [1.075, 2.086]	0.7018

Appendix

Table 41

Subset	n	Stage	Pooled mIoU [95% BCa CI]	BIoU
all	400	initial	0.2763 [0.2653, 0.2885]	0.1152
all	400	protected repair	0.3159 [0.3028, 0.3295]	0.1515
all	400	cleanup	0.3596 [0.3449, 0.3740]	0.1858
repair triggered	263	initial	0.2410 [0.2315, 0.2519]	0.0849
repair triggered	263	protected repair	0.2984 [0.2833, 0.3144]	0.1402
repair triggered	263	cleanup	0.3384 [0.3213, 0.3553]	0.1708

Appendix

Table 42

Fig. C.4: Stage-detail audit: (a) paired plan-mean repair-minus-initial mIoU differences with 95% BCa intervals; (b) class IoU under the frozen final transform. Repair-skipped images are identical before cleanup. Frame-conditioned results and group counts are in Table C.16 .

Method	Band	Errors	Boundary	Omitted	Added	Confused
SALI-FP	0.25%	56,350,103	19,719,198 (34.99%)	12,199,963 (21.65%)	17,045,832 (30.25%)	7,385,110 (13.11%)
SALI-FP	0.5%	56,350,103	25,050,419 (44.45%)	10,094,113 (17.91%)	16,115,423 (28.60%)	5,090,148 (9.03%)
SALI-FP	1%	56,350,103	31,796,315 (56.43%)	7,180,240 (12.74%)	14,375,243 (25.51%)	2,998,305 (5.32%)
CubiCasa5K / argmax	0.25%	20,259,975	7,341,614 (36.24%)	4,223,581 (20.85%)	7,850,721 (38.75%)	844,059 (4.17%)
CubiCasa5K / argmax	0.5%	20,259,975	8,266,735 (40.80%)	3,842,951 (18.97%)	7,608,690 (37.56%)	541,599 (2.67%)
CubiCasa5K / argmax	1%	20,259,975	9,604,690 (47.41%)	3,222,779 (15.91%)	7,140,895 (35.25%)	291,611 (1.44%)

Appendix

Table 44

Geometric cause	Original objects	Delivered objects
empty	7765	7765
hole or ring structure	4	0
malformed	6	6
other invalid	90	11
self intersection	101478	22935
valid	673880	752510

Appendix

Table 45

Fig. C.5: ArchP10k representation sizes per plan. Boxes span the interquartile range, center lines mark medians, and whiskers extend to the furthest observations within 1.5 IQR. Outliers are omitted only from this display; all plans remain in the statistics. The count axis is logarithmic; related export statistics are in Table C.6 .

Local method	n	Median (s)	IQR (s)
CubiCasa5K / argmax	400	0.380	0.368–0.388
CubiCasa5K / polygons	400	6.804	6.589–7.301
MiT-UNet	400	0.050	0.049–0.052
Raster2Seq	400	1.153	0.866–1.728

Appendix

Table 47

Class	Objects	Self-cross	Empty	Other invalid	Invalid share
balcony	15,524	296	160	1	2.94%
bathroom	32,109	261	11	0	0.85%
bedroom	49,056	2,839	1,582	12	9.04%
corridor	22,345	2,044	779	8	12.67%
door	160,745	216	148	1	0.23%
elevator	10,389	104	63	0	1.61%

Appendix

Table 48

Source category	Objects	Self-cross	Empty	Other invalid	Invalid share
residential other	440,079	13,684	4,867	47	4.23%
office	51,837	1,417	446	7	3.61%
interior general	38,446	1,218	429	7	4.30%
hotel	38,203	1,317	308	7	4.27%
education	52,371	1,200	317	5	2.91%
cultural other	45,099	966	387	6	3.01%

Appendix

Table 49

Method	GT class	GT pixels	Boundary error	Interior omitted	Interior confused
SALI-FP	Door/opening	2,162,846	65.99%	1.38%	4.23%
SALI-FP	Window	4,708,442	55.19%	4.48%	7.66%
CubiCasa5K / argmax	Door/opening	2,162,846	29.63%	0.08%	1.68%
CubiCasa5K / argmax	Window	4,708,442	18.84%	0.17%	1.37%
CubiCasa5K / polygons	Door/opening	2,162,846	30.62%	0.70%	2.44%
CubiCasa5K / polygons	Window	4,708,442	25.18%	2.70%	4.10%

Appendix

Table 50

Scope	Plans	Method	mIoU	Added pixels
complete test	400	SALI-FP	0.3596	16,115,423
complete test	400	CubiCasa5K / argmax	0.7390	7,608,690
complete test	400	CubiCasa5K / polygons	0.6612	6,912,552
nonempty room reference	398	SALI-FP	0.3603	15,362,301
nonempty room reference	398	CubiCasa5K / argmax	0.7409	6,804,133
nonempty room reference	398	CubiCasa5K / polygons	0.6628	6,081,348

Appendix

Table 51

Method	Four-class mIoU [95% BCa CI]	Room F1	Corner F1	Angle F1
SALI-FP	0.3596 [0.3449, 0.3740]	0.445	0.159	0.093
CubiCasa5K / argmax	0.7390 [0.7270, 0.7496]	N/A	N/A	N/A
CubiCasa5K / polygons	0.6612 [0.6422, 0.6766]	0.630	0.327	0.236
Raster2Seq	N/A	0.763	0.546	0.368

Appendix

Table 52

Fig. C.6: CubiCasa5K benchmark calibration: (a) GT mIoU; (b) native room, corner, and angle F1 with 95% BCa intervals. The lower band reports native room medians; IQRs are in Table C.15 . Fixed-tolerance curves are in Fig. C.3.

Stage	mIoU [95% BCa CI]	BIoU	Boundary F1
Initial $S_{0}$	0.2099 [0.2024, 0.2193]	0.0642	0.2358
Protected $S_{1}$	0.2015 [0.1942, 0.2101]	0.0610	0.2260
Cleanup labels	0.2074 [0.2000, 0.2162]	0.0570	0.2100
Final representation	0.3596 [0.3449, 0.3740]	0.1858	0.4962

Appendix

Table 54

Fig. C.7: Recorded stages in original and frozen final coordinates: pooled mIoU and boundary IoU with 95% BCa intervals. The band reports plan-mean repair-minus-initial differences; complete paired intervals and class trajectories are in Fig. C.4.

Fig. C.8: Systematic pixel-error diagnostics: SALI-FP, CubiCasa5K argmax, and CubiCasa5K polygons, with GT-row-normalized confusion and the disjoint 0.5%-diagonal error decomposition.

Factor	Representation	rho [95% CI]	Holm p
ink ratio	auxiliary rcr	0.072 [0.050, 0.094]	0.0012
ink ratio	delivered rcr	0.030 [0.009, 0.051]	0.0141
edge density	auxiliary rcr	-0.093 [-0.114, -0.071]	0.0012
edge density	delivered rcr	-0.101 [-0.120, -0.080]	0.0012
component count	auxiliary rcr	0.004 [-0.018, 0.027]	1.0000
component count	delivered rcr	0.002 [-0.019, 0.023]	1.0000

Appendix

Table 57

Outcome	Hard mean	Non-hard mean	Difference [95% CI]	Holm p
auxiliary rcr	0.9507	0.9582	-0.0075 [-0.0102, -0.0047]	0.0003
delivered rcr	0.9133	0.9324	-0.0191 [-0.0269, -0.0120]	0.0003
protected	0.2275	0.1729	0.0546 [0.0043, 0.1099]	0.0436

Appendix

Table 58

Resource	Located material	Use in this study
CubiCasa5K [ 2 ]	Official images, SVG labels, code, trained checkpoint	400 GT reruns; full-channel polygon rerun; 30 retained illustrations
MiT-UNet [ 25 ]	Official code, trained weights, regional-data link	400 visible-wall tests; 30 illustrations
Raster2Seq [ 27 ]	Official code and checkpoint	400 CPU native-polygon inferences; room evaluation on 398 reference cases
FloorPlanFormer [ 28 ]	Training code; validation/test links	No official inference checkpoint located
R2V / R3D [ 3 , 5 ]	Code and test lists; incomplete authorized original inputs	No SALI-FP benchmark values
MLStructFP [ 9 ]	Official dataset documentation and request process	Not downloaded as an unrestricted mirror

Appendix

Table 59

Item	Prepared material	Required completion
ArchP10k accuracy	200 category/complexity-stratified plans; two blank annotation templates	Independent annotations, adjudication, frozen GT metrics
Gate correctness	200 accepted and 200 rejected proposals, blinded	Independent correctness scores; weighted estimates and agreement
Causal stage ablation	Recorded-stage paired evaluation	Controlled reruns holding models, prompts, and inputs fixed
Pretraining-exposure diagnostics	Exposure is not auditable from deployment aliases	Frozen near-duplicate and memorization probes with control images; no absence-of-contamination claim
Open-weight substitution	Interface and label protocol	Frozen image/VL weights; matched-subset inference and GT evaluation
Independent repeatability	Case and prompt identifiers	Multiple uncached model runs; variance and output-identity audit

Appendix

Table 60

Explore similar work

ARCH-B: Architectural Representation, Comprehension and Hierarchy Benchmark
Sep 28, 2026Kieran Sagar Parikh, Jose Luis Garcia del Castillo y LopezSemantic Correspondence Cross-Modal Alignment
Harvard University · Northeastern University
Tokenization Allows Multimodal Large Language Models to Understand, Generate and Edit Architectural Floor Plans
Mar 12, 2026Sizhong Qin, Ramon Elias Weber, Xinzheng LuMultimodal Generation Multimodal Large Language Models
Tsinghua University · UC Berkeley
Multimodal Floorplan Encoding: Learning Dense Modality-Invariant Representations
Sep 11, 2026Xavier Anadón, Rémi Pautrat, Rui WangRepresentation Learning Cross-Modal Representation Learning
University of Zaragoza, Spain · Microsoft, Switzerland

Sep 28, 2026J/K move · Enter open · S save

ARCH-B: Architectural Representation, Comprehension and Hierarchy Benchmark

Kieran Sagar Parikh, Jose Luis Garcia del Castillo y Lopez

Harvard University · Northeastern University

Semantic Correspondence Cross-Modal Alignment Cross-Modal Feature Matching Multimodal Model Evaluation

Open paper AskPDF arXiv

Multimodal models increasingly interpret visual environments, but their ability to recognize the same building across photographs, floor plans, elevations, sections, and renderings remains poorly characterized. We introduce ARCH-B, a benchmark of 354 four-choice questions across 11 cross-representational archetypes, constructed from a building-linked corpus of 3.9 million architectural images using visually similar distractors, model-guided difficulty screening, and manual validation. We evaluate 25 multimodal models and collect 5,830 responses from non-expert human participants. Model accuracy ranges from 10.45% to 83.90%, compared with a human baseline of 35.35%. Models perform comparatively well on mixed-representation outlier detection and photograph matching, but remain weaker on floorplan-to-photograph correspondence. Human and model difficulty across archetypes is only weakly correlated (Spearman's (ρ=0.33)). Held-out evaluation confirms that the difficulty identified during screening generalizes beyond the curation models. ARCH-B provides a diagnostic evaluation of visual correspondence and representation transfer across architectural media.

Abstract

Explore similar work

ARCH-B: Architectural Representation, Comprehension and Hierarchy Benchmark

Tokenization Allows Multimodal Large Language Models to Understand, Generate and Edit Architectural Floor Plans

Multimodal Floorplan Encoding: Learning Dense Modality-Invariant Representations