PaperlayerView exact revision on arXiv

Source-linked AI summary

When Should a Network Emit Geometry, and When Should It Detect It? Readout, Reconciliation, and Representation in Floorplan Vectorization

He Zhang

arXiv:2608.25608v1cs.CV

TL;DR

Floorplan vectorization can read geometry by autoregressive emission or dense detection, but evidence is limited on when each readout and representation works best. The paper compares these choices on matched systems, adds human-edit cost and benchmark resources, and finds that readout depends on domain and plan size, representation is secondary under controlled comparison, and output fusion helps while input conditioning does not.

  • Problem

    Existing floorplan vectorization evidence does not isolate readout from architecture and training, and standard F1 or IoU misses the human effort required to correct drafts.

  • Method

    The paper compares sequence emission and heatmap-based graph detection on one trained network, evaluates room-centric reconciliation against wall-first emission, and introduces edit-cost evaluation and benchmark resources.

  • Results

    Detection leads on large real scans, sequence decoding leads on covered clean-render styles, room-centric reconciliation matches or exceeds wall-first wall quality, and output fusion raises wall F1 by 7 points while input conditioning gives no gain.

  • Takeaways & Limitations

    The preferred readout depends on plan size and domain, while practical complementarity is better exploited by deterministic output fusion than by conditioning one model on another’s draft.

  • Takeaways & Limitations

    The controlled representation study compares systems of unequal maturity, leaves training supervision unisolated, and makes no claim about rooms or openings.

Abstract

from arXiv · show

A network trained to recover the walls, openings, and rooms of a rasterized floorplan can produce its output in two ways: by emitting the geometry as an autoregressive coordinate sequence, or by detecting it on dense junction and centerline heatmaps and assembling a graph. We compare the two readouts on the same trained network. On real scans (CubiCasa5K) detection is better on every wall measure (+2.7 wall F1 at tolerance 0.05, +5.1 at 0.015; paired bootstrap intervals exclude zero), and reading an opening heatmap the decoder never used raises opening F1 by 2.6x without retraining. Within real scans the readout's advantage grows with plan size and reverses on small plans; on clean vector renders sequence decoding is better by 5 to 8 points where its training covered the render style, while under full domain shift the readout, given calibrated thresholds, stays ahead; neither ink density nor plan size explains the reversal. With matched data and recipe, a room-centric system with a reconciliation step and a wall-first sequence model reach comparable wall quality, so the output representation matters less than is usually assumed. A prior from the other family helps at the output but not at the input: deterministic fusion of the two outputs raises wall F1 by 7 points, whereas conditioning one model on the other's output gives no gain in three forms, including two ground-truth-content controls. We also provide an edit-cost metric that scores a draft by the human work needed to correct it, corrected CubiCasa5K annotations, and ResPlan-FP, a CC BY 4.0 benchmark of 16,998 plans with frozen splits and three baseline tracks. Code, the benchmark, and the corrected annotations are available at https://github.com/Cyprinus12138/fpvec-lab

1 Introduction

The paper separates geometry readout from output representation, comparing emission and detection on matched systems and data. Detection leads on large real scans, decoding leads on covered clean-render styles, representation is secondary under controlled comparison, and output fusion is more useful than input conditioning.

  • Resources: The paper introduces an edit-cost metric, corrected CubiCasa5K annotations, and ResPlan-FP, a CC BY 4.0 benchmark with 16,998 plans.The benchmark has frozen splits and three baseline tracks.
  • Readout: 2.7 wall F1 points at tolerance 0.05 and 5.1 at 0.015 favor graph detection over sequence emission on CubiCasa5K scans.Paired bootstrap 95% intervals are [1.4, 4.0] and [3.7, 6.4].
  • Readout: Sequence decoding leads by 5 to 8 points on clean renders whose style it covered, while calibrated detection stays ahead under full domain shift.Within real scans, detection’s advantage grows with plan size and reverses on small plans; ink density and plan size do not explain the cross-domain reversal.
  • Readout: Opening F1 rises from 0.25 to 0.64 when the decoder reads its jointly trained opening heatmap, without retraining.
  • Representation: With matched data and recipe, room-centric emission plus reconciliation reaches comparable or better wall structure than native wall-first emission.The controlled comparison attributes earlier failures to ground-truth convention, reconciliation, and domain shift rather than representation alone.
  • Complementarity: 7 wall F1 points result from deterministic output fusion, whereas conditioning on another system’s output provides no gain across three input forms.Two controls supplying ground-truth content moved the score by at most 0.3 points.

3 Task, Systems, and Metrics

The task is raster-to-vector reconstruction of watertight floorplans, evaluated with geometric, topological, and human-correction measures. The comparison treats room-centric and wall-first systems as complete systems because reconciliation and post-processing affect their outputs.

  • Task: The output is a watertight vector plan containing wall centerline segments, door and window intervals, and labeled room polygons.
  • Systems: Room-centric Raster2Seq emits labeled room polygons, then reconciliation welds vertices, deduplicates shared edges, and T-splits them to derive walls.
  • Systems: Wall-first emission produces walls, room cycles with explicit open edges, and openings, while snapping dense predictions to obtain watertightness.Near-coincident wall segments can still require snapping or merging.
  • Metrics: F1 uses Hungarian matching at IoU≥0.5 across tolerances 0.015–0.08, alongside watertightness and double-wall counts.Watertightness is interpreted together with duplicate-wall counts because duplicated walls can close regions trivially.
  • Metrics: Edit cost models correction effort by matching geometry and assigning weighted MOVE, TYPE, CREATE, DELETE, and CONVERT operations.The edit script is determined by matching rather than minimized over all possible scripts.
  • Metrics: The operation weights follow editor-interaction ordering, but their magnitudes are hand-set and calibration against deployed-editor logs remains future work.System rankings remain unchanged throughout the weight region satisfying Eq. (2).

4 The Wall-First Network and Its Two Readouts

A shared encoder supports two wall-first readouts: autoregressive geometry emission and parameter-free graph detection from dense heatmaps. The study also compares these readouts with a separate room-centric system using reconciliation.

  • Architecture: The network reuses a Raster2Seq backbone and grammar-masked decoder with 32-bin coordinate quantization per axis.The grammar restricts valid token types and prevents duplicate wall IDs or references to un emitted walls, but not near-coincident segments.
  • Architecture: Three jointly trained heatmaps predict junctions, wall centerlines, and openings at 128 × 128 resolution with sub-pixel offsets.
  • Readouts: Emit decodes walls, rooms, and openings, then snaps wall endpoints to junction peaks; detect forms graph walls from junction peaks and centerline coverage without trained parameters.Detect uses graph faces as rooms and calibrated thresholds.
  • Experimental setup: The walls-only arm is the readout-comparison network, conditioning-arm base, and fusion donor; both arms share training recipes across data regimes.
  • Comparison system: Raster2Seq is a separate room-centric system whose reconciliation derives walls by welding vertices, deduplicating shared edges, and T-splitting.

5 Experimental Setup

Experiments isolate readout, representation, and fusion effects in floorplan vectorization. Detection is strongest on large real scans, while reconciliation and output-level fusion materially improve practical quality.

  • Emit versus detect: Detection exceeds sequence decoding on real scans by 2.7pp wall F1 at tolerance 0.05 and 5.1pp at 0.015.The paired bootstrap intervals exclude zero on 296 CubiCasa5K plans.
  • Emit versus detect: Readout performance reverses with plan size: it trails by 3.7pp on 5–22-wall plans but leads by 8.7pp on 32–70-wall plans.The empirical rule is to decode small real plans and detect large ones.
  • Emit versus detect: Sequence decoding leads by 5 to 8 points on clean renders when training or calibration covers the render domain, and neither ink density nor plan size explains the reversal.Under domain shift, calibrated readout thresholds preserve the advantage.
  • Emit versus detect: Reading the unused opening heatmap raises opening F1 from 0.246 to 0.644 with a family heuristic, or 0.799 when the family is ignored, without retraining.The readout alone yields a 2.6× improvement under the internal CubiCasa5K matcher.
  • Representation is not the main factor: With matched data and recipe, room-centric output after reconciliation is not behind this wall-first implementation on wall structure, but its coverage excludes non-room-bounding walls.Room-centric recall is higher on boundary walls; the aggregate gap comes from dangling, stub, and exterior walls that room polygons cannot express.
  • Output-level fusion: Output-level fusion reaches 0.853 wall F1 at tolerance 0.05 and reduces edit cost, whereas conditioning on another output fails in all three tested forms.Fusion gains 7.1pp over the room-centric system, while ground-truth-content controls move conditioning scores by at most 0.3 points.

7 Limitations

The controlled comparison makes system-level claims rather than isolating output representations, and several study limitations constrain interpretation.

  • Scope: The room-centric and wall-first comparison evaluates complete systems, not representations in isolation.The study therefore does not independently vary representation while holding every other system component fixed.
  • Scope: Room-centric outputs and openings are under-tuned, so the study bounds room emission’s wall-structure cost but says nothing about rooms or openings.The room-centric arm continues a published released checkpoint, while the wall-first arm is the authors’ system.
  • Metric: Edit-cost magnitudes are hand-set; although operation ordering survives a weight sweep, calibration remains future work.This limits how directly the reported costs should be interpreted as human correction effort.
  • Readout generalization: The cross-domain reversal appears only when decoder training covered the render style, while its presumed style-prior mechanism is not tested.Under full domain shift, calibrated detection remains ahead; neither ink density nor plan size explains the reversal.
  • Interpretation: No experiment varies training supervision alone, so earlier room-centric failures reflect evaluation convention, reconciliation, and domain shift jointly.The paper does not identify a single causal source for those failures.

8 Conclusion

Performance is driven less by output representation than often assumed: matched room-centric and wall-first systems have nearly equal wall quality, while readout choice and output fusion matter substantially.

  • Representation: With matched data and recipe, room-centric and wall-first systems reach almost the same wall quality.Remaining differences mainly reflect representational envelope, evaluation convention, and post-processing or reconciliation.
  • Complementarity: 7 wall-F1 points come from fusing complementary outputs, whereas using one system’s output as an input prior provides no gain.Input conditioning can cost 2–3 points when the draft contains errors.
  • Readout: Dense detection outperforms autoregressive emission on large real scans, while decoding wins on small plans and style-matched clean renders.Under full domain shift, calibrated detection stays ahead, and the reversal is not explained by ink density or plan size.
  • Resources: The paper releases an edit-cost metric, corrected CubiCasa5K annotations, and the redistributable ResPlan-FP benchmark.These resources accompany the findings on readout, representation, and cross-system complementarity.

A Edit-cost metric: procedure and breakdown

The edit-cost metric converts predicted-versus-ground-truth discrepancies into weighted correction operations, averages their total per plan, and reports operation counts in a CubiCasa5K breakdown.

  • Wall matching: Predicted and ground-truth walls are Hungarian-matched within a move radius of 4t, with t = 0.05 of the plan frame.A nearby prediction is treated as a displaced wall rather than a delete followed by a create.
  • Wall operations: Matched wall pairs incur MOVE costs proportional to distance in units of t and a TYPE operation when wall types differ.Wall types are exterior, interior, partition, or shelter.
  • Unmatched walls: Unmatched ground-truth walls count as CREATE operations, while unmatched predictions count as CONVERT over uncovered open edges or DELETE otherwise.A double wall therefore incurs one DELETE, whereas a missing real boundary is charged as CREATE.
  • Openings: Openings are matched separately for doors and windows; matched pairs incur OPEN-MOVE weighted 0.5 per unit of t, and unmatched openings cost 4 each.The unmatched cases are OPEN-CREATE and OPEN-DELETE.
  • Aggregation and breakdown: The reported edit cost is the per-plan weighted total averaged over the test set, with Table 11 breaking down mean operations for 287 official CubiCasa5K test plans.Its wall-first total is 110.2, differing from Table 4’s 113.9 because the roster differs.

B Synthetic data generator

The synthetic generator creates verified raster plans, aligned vector ground truth, and millimeter-space scenes from deterministic procedural layouts whose unit statistics are fitted to real listings.

  • Generation: Each generated sample returns a rendered raster plan, aligned fpeval vector ground truth, and a millimeter-space scene description from a random seed.The same scene can be re-rendered in another style, and generation is deterministic given the seed.
  • Unit statistics: Unit descriptions are fitted to real residential listings using 18,203 parsed bedroom-and-bathroom counts from 29,300 Singapore property titles.The resulting joint distribution and area bands are summarized in Table 12.
  • Scope: Synthetic drawings are not claimed to resemble any particular real drawing style, and evaluation sets come from other markets.This is an explicit scope boundary for interpreting synthetic-to-real similarity.
  • Layout: A public/private room-program tree fixes door topology, while area budgeting can drop optional rooms and rectangular envelopes use aspect ratios from 1.0 to 2.2.The layout is placed on a 10 mm grid.
  • Openings and validity: Doors follow the access tree, windows appear only on exterior walls of habitable rooms, and unreachable-room layouts are rejected.Furniture is placed by rejection sampling.
  • Rendering: Plans use four style families—clean CAD, marketing, brochure, and hand-drawn—with geometry-preserving raster rendering and degradation operations.Wall bands are buffered centerlines with opening gaps subtracted, while symbols represent doors and windows.
  • Quality control: Every sample passes watertightness and reachability verification before writing, with failed samples regenerated and metadata recorded in the manifest.The manifest stores seed, unit description, style, degradation operations, transform, and verification result.

C Evaluation definitions and the reconciliation step

Evaluation matches walls, openings, and rooms under explicit geometric criteria, while reconciliation converts room polygons into a wall graph for comparison.

  • Evaluation definitions: Walls are matched one-to-one using angle, sampled mean distance, and axis-overlap constraints, with Hungarian matching assigning minimum-cost candidate pairs.
  • Evaluation definitions: Opening centers are matched separately by door and window family within 2t, using a wider tolerance because centers are less precisely defined than wall endpoints.
  • Evaluation definitions: Room polygons are matched one-to-one at IoU≥0.5, with graph-based systems deriving rooms as faces of their wall graph.
  • Evaluation definitions: rval measures the fraction of plans whose room polygons are simple, non-overlapping, and tile the exterior interior, while allowing open-plan boundary edges.
  • Evaluation definitions: An ink gate retains candidate walls with at least 0.5 dark-pixel coverage, using perpendicular ±3 px sampling to accommodate hollow and double-line styles.
  • Reconciliation step: Reconciliation unpads and rescales room polygons, then welds their vertices so two sides of a shared wall become a unified wall representation.

D ResPlan-FP construction

ResPlan-FP converts licensed vector floorplans into centerline-wall supervision and releases a reproducible benchmark with fixed splits and evaluation tracks.

  • Dataset construction: 17,000 residential plans supply vector walls, openings, room polygons, and labels under CC BY 4.0 licensing.
  • Dataset construction: Wall bodies and opening bodies are united, rasterized, morphologically closed, skeletonized, and traced into a centerline wall graph.
  • Dataset construction: Graph paths are simplified into straight segments, welded at junctions, snapped to Manhattan axes, and paired with openings projected onto centerlines.
  • Dataset construction: Plans render as clean 256 px line drawings with wall bands, door swings, jamb lines, and no furniture or text.
  • Benchmark protocol: 16,998 accepted plans are split 14,998/1,000/1,000 using seed 42, with archived test identifiers and no near-duplicate removal.
  • Benchmark protocol: The benchmark defines zero-shot, finetuned, and LLM zero-shot evaluation tracks.

E Readout calibration

Graph-readout calibration selects thresholds and chaining settings on validation data for each evaluation domain, then freezes them before testing.

  • Calibration procedure: The graph readout tunes junction and centerline thresholds, chord coverage, and collinear-edge chaining on the validation split before test evaluation.
  • CubiCasa5K: CubiCasa5K freezes θJ = 0.55, θC = 0.4, κ = 0.5, and no chaining after validation-grid expansion.
  • CubiCasa5K: 22pp of wall F1 is lost by chaining collinear edges at t = 0.05 on CubiCasa5K validation because ground-truth walls are split at junctions.
  • ResPlan-FP: Extending ResPlan-FP calibration yields validation wall F1 0.913 and test wall F1 0.915 at t = 0.05, versus 0.896 at t = 0.015.
  • ResPlan-FP: The frozen CubiCasa5K configuration transfers almost as well to ResPlan-FP, reaching test wall F1 0.914 at t = 0.05.
  • Open-plan tier: Open-plan calibration freezes θJ = 0.55, θC = 0.35, κ = 0.5, and no chaining after an in-domain 59-plan calibration set.

F Training details

Training compares a wall-first network with room-centric Raster2Seq arms under matched data regimes, using corrected CubiCasa5K supervision and shared synthetic mixtures.

  • Model training: The wall-first network combines a Raster2Seq backbone and causal decoder with dense junction, centerline, and opening heatmap predictions.
  • Model training: Raster2Seq room-centric arms use polygon-sequence objectives, 12 semantic classes, 512-token sequences, 32 coordinate bins, and EMA evaluation weights.
  • Data regimes: The cc5k+synth mix pairs 3,328 official plans with 3,327 synthetic plans, while synth-only uses synthetic plans alone.
  • Annotation correction: Corrected CubiCasa5K annotations rebuild rooms as faces of a welded wall graph and project openings onto wall centerlines after conversion failures were identified.

G LLM zero-shot protocol

The zero-shot protocol evaluates Gemini 3.1 Pro on 200 benchmark plans using a constrained JSON vectorization prompt and fixed image preprocessing. Parsed outputs receive wall, room, opening, rval, and cost scores, while fusion results use separate decodability rosters.

  • Model and input: Gemini 3.1 Pro receives one 256 px benchmark render per plan, upscaled to a 1024 px longer edge and sent as PNG.The API configuration uses temperature 0, a 16,000-token output budget, and no retries or self-correction.
  • Prompt and output: The requested output is a single JSON object containing walls with endpoints, thickness, curvature, openings, and rooms linked to wall identifiers.The response is parsed into the fpeval plan format, and an unparseable object scores zero.
  • Prompt and output: The prompt requires tracing structural walls and openings while ignoring furniture, fixtures, text, and colour fills.Coordinates are normalized with the longer image edge set to 1024.
  • Evaluation: 0.811 wall F1 at t = 0.05 and 0.472 at 0.015 are reported alongside rooms 0.708, openings 0.752, rval = 0.620, and total cost of about US$6.These results use a seeded random subset of 200 plans from the 1,000-plan test split.
  • Fusion evaluation: 0.855 wall score at t = 0.05 is reached by cross-representation fusion on 287 of 296 plans where both source models decode.The stated base scores are 0.786 for r2s_mix and 0.793 for wf_mix; fusion gains 6.9pp over r2s_mix.
Loading 2608.25608v1…