acceptodds
Under review as a conference paper at ICLR 2027

CartoGenS Geographically ConditionedVector Scene Generation forBuilding Generalization

Abstract

Building generalization across map scales requires coordinated changes in building shape, count, and boundary structure, altering both geometry and topology of the vector scene. These changes cannot be decided building by building, yet existing learning methods process individual buildings, local groups, or rasters that require vectorization. We introduce CartoGenS, an end-to-end framework that generates coarser-scale vector building scenes without predefined source-to-target building correspondences. A shared encoder represents building boundaries, road context, and source-scene physical extent. A single set of learned queries predicts vertices for the entire target scene, and directed successor prediction links them into cycles that form exterior and interior rings, so the number and nesting of output polygons are predicted rather than inherited from the source buildings. Successor scores include a learned relative-geometry term. For scene-level training and evaluation, we construct paired vector building scenes with road context for a fixed 1:10,000-to-1:25,000 transition. CartoGenS achieves the highest overall polygon IoU, panoptic quality, and boundary IoU among the evaluated methods, exceeding the strongest baseline by 9.48 percentage points in boundary IoU. Full factorial ablations highlight the importance of geographic conditioning, and relative-geometry scoring further improves overall performance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
vector scene generation, building generalization, geographic conditioning, geometric deep learning, structured prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QwQ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper foucus on building generalization (1:10k → 1:25k) as scene-level vector generation. Building vertices, road tokens, and a scene-scale scalar are encoded jointly. DETR-style queries predict target vertices, and a PolyWorld-style successor module links them into cycles that form exterior and interior rings, so building count and holes are predicted rather than inherited. On 35,811 OS Open Zoomstack scenes with a geographic split, the method achieves the best polygon IoU, PQ, and boundary IoU against raster, LLM, and ArcGIS baselines.

Critical strengths

Well-motivated problem formulation. Limitations are stated candidly, and the dataset could become a useful benchmark if it can be made public.

Critical weaknesses
  1. The representation, successor module, and loss come from PolyWorld. The additions (vector input, DETR queries, a road flag, a scale scalar, a distance prior) are modest. PolyWorld (Pix2Poly, and RoomFormer) are not compared, although they could be trained on the same rasterized masks used for the raster baselines. Such a comparison would isolate the actual contribution.
  2. The no-road ablations look like optimization failure. Without roads, 60.9% of outputs are empty and PQ is 0.080, while road-free raster baselines reach 0.38–0.53. The road effect may therefore reflect training dynamics rather than geographic information. Dummy-token controls and multiple seeds are needed.
  3. The raster baselines receive neither roads nor a fixed scale, yet Base + road (PQ 0.512) already trails the road-free RoofX-Net (0.527). ArcGIS thresholds are untuned, raster outputs lack standard post-processing, and dedicated generalization models are missing. The LLM baseline, with no training at all, is only 2.9 PQ points behind.
  4. No validity or cartographic checks. Rings are repaired with buffer(0), and the paper does not report self-intersection rates, orphan holes, minimum separation, or road conflicts. Full also returns empty output for 6.35% of the simplest scenes.
Questions for authors
  1. How do PolyWorld and RoomFormer perform on the same rasterized input with a road channel?
  2. Does replacing roads with dummy tokens reduce Base’s empty-output rate to a similar degree?
  3. Does the main ranking hold when raster baselines receive roads and a fixed resolution?
  4. What fraction of Full’s empty outputs stems from orientation errors?
  5. How was the 1:25k layer produced, and how does a tuned ArcGIS workflow perform?
  6. I notice that for qulitative results on Figure 6: Road-context ablation, your method shows sharp edge, what causes that?
2 comments
AuthorsOpenReview ↗

Response to Reviewer QwQ3

We thank Reviewer QwQ3 for the comments. Point-by-point replies below.

W1 / Q1. PolyWorld and others on raster input. We trained PolyWorld from scratch on the same building + road rasters, with the official architecture and the original multi-stage schedule. Test results:

MethodPQPoly. IoUBIoURaw validitySelf-int.Empty
CartoGenS (paper)0.6160.6320.66170.4%10.3%3.9%
PolyWorld, buildings + roads†0.5760.6280.59169.9%21.4%2.9%
PolyWorld, buildings only†0.0190.0160.0123.1%21.3%95.1%

We could not train RoomFormer and Pix2Poly within the response period.

W2 / Q2. Is the no-road ablation an optimization failure; dummy tokens; seeds. Dummy-token control (same token count, road coordinates set to zero), test results:

BaseBase + dummyBase + roads
Empty60.9%48.7%19.4%
PQ0.0800.1160.512

Dummy tokens recover only 8% of the PQ gain from real roads; most of the effect comes from road geometry. An empty output is a degenerate solution of the successor objective, where every query points to itself. All results use one fixed seed.

W3 / Q3. Roads and fixed resolution for raster baselines; post-processing; gap to the LLM. Raster baselines with roads, test results:

MethodPQPoly. IoUBIoUTFD↓VCR
CartoGenS (paper)0.6160.6320.6610.5091.31
RoofX-Net → + road†0.527 → 0.6680.570 → 0.6980.396 → 0.6240.631 → 0.542≈31
MSEONet → + road†0.492 → 0.6340.544 → 0.6700.373 → 0.5710.662 → 0.567≈31
SAM-BG → + road†0.380 → 0.4850.568 → 0.6760.382 → 0.5740.684 → 0.593≈33

Roads raise PQ by 0.10–0.14 for all three models, matching CartoGenS's large drop without roads (0.512 → 0.080): whatever the architecture, geographic context such as roads is necessary. With roads, RoofX-Net is ahead in PQ and polygon IoU and CartoGenS on boundaries and shape; our comparison conclusion is adjusted accordingly. We could not finish the fixed-resolution raster variants or simplification post-processing within the response period; the raster outputs here are unsimplified. As for the LLM baseline, it outputs fewer vertices than the reference (VCR 0.87) and has lower BIoU (0.566 vs. 0.661).

W4 / Q4. Validity, empty outputs and orientation. We added an analysis of the unrepaired output: the paper model has 70.4% raw validity, 10.3% self-intersections and 3.9% empty outputs. Per scene: 16.4% have a self-intersecting ring, 10.1% overlapping exterior rings, 4.4% a hole outside every exterior ring. Self-intersections have two main causes. First, Hungarian matching uses coordinates only; when two queries land on the same vertex, the label assignment is close to random, both can end up in the ring at inference, and rings with such a duplicate self-intersect 5× as often. Second, 46.6% of reference vertices are densified arc points 0.3–0.4 m apart, below the coordinate error; the order of neighbouring points is flipped on 19.4% of arc edges vs. 2.4% of other edges.

Other methods self-intersect less partly because raster polygons are traced from a mask and cannot self-intersect, and partly because they mostly do not represent arcs. Share of target arc points reproduced within 0.5 m: CartoGenS 65.7%, RoofX-Net + road 18.0% (staircase tracing, ~31× the vertices), GPT 12.6%, PolyWorld 6.8%, ArcGIS 0.5%. PolyWorld, also vertex-and-edge based, has 21.4% self-intersections.

Of the 3.86% empty outputs, 3.73% are scenes where no ring is decoded (all queries self-loop) and 0.12% are scenes where all rings are clockwise and read as holes, so orientation accounts for 3% of empties. The 6.35% empty rate on single-building scenes is the same all-self-loop failure. We have improved the method to address these causes. Minimum-separation and road-conflict checks are not added yet.

Q5. The 1:25,000 layer. It is the district building layer published in OS Open Zoomstack, produced by Ordnance Survey and used as published. ArcGIS tuned on validation (54 settings, one test run): PQ 0.493 vs. 0.616 for CartoGenS.

Q6. Sharp edges in Fig. 6. These are spikes from duplicate vertices connected in the wrong order, the first cause in W4: two queries sit on one corner, and the ring visits one, leaves, then comes back to the other. We have improved the method to address this cause.

Thanks for the author's reply. After reading the rebuttal, I still have some concerns. (1) with roads, the raster baselines match or exceed CartoGens on overlap and instance metrics, and PolyWorld is close as well, the abstracts's claim is therefore overstated. (2) The road ablation may largely reflect escaping a degenerate solution rather than geographic information. (3) "We have improved the method to address this cause." doesn't provide details about how and to what extent. Therefore, I will keep my score.

Reviewer_9oc3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces CartoGenS, a model that generates building polygons for road bounded blocks. Unlike prior approaches that simplify infividual buildings or aggregate groups, CartoGenS predicts a pool of vertices shared across the entire scene. Instead of conditioning on number of holes or buildings it is predicted. A paired benchmark is introduced and comparisons are made against raster networks, and other approaches. CartoGenS generally performs best on this benchmark.

Critical strengths

Scene level problem formulation: Moving conditions from the building level to scene level predictions is a good generalization to make the model more flexible to realistic scenarios.

New benchmark and evaluation: The increased complication to the benchmark in comparison to prior work is a good step towards a more realistic evaluation. Separating area agreement from instance recovery is a principled way to evaluate the model.

Critical weaknesses

Incomplete baselines:

(a) The raster baselines are given building masks only, while CartoGenS and other methods are given road information (Sec 4.1). The ablation in Table 4 suggests that road context is important, so it's hard to disentangle the impact of road conditioning. Giving the raster networks a road channel would make the comparison more convincing.

(b) The only non-learned baseline is ArcGIS Pro with a single aggregation distance and simplification tolerance for every scene (Sec 4.2). Much of the the test set seems to be single-building blocks without holes (Table 1), where the target is close to a cleaned-up union of the inputs. Comparing against a tuned algorithmic approach and a trivial baseline would show the necessity of the learned approach.

Ablation evidence: Without road tokens the model returns nothing for a majority of scenes, and for 73% of single-building blocks without holes (Sec 4.3). Not sure if this suggests training convergence issues or that roads carry useful context.

Questions for authors
  1. Does gijg_{ij} act as a learned edge-length penalty? Since the Δxij\Delta x_{ij}, Δyij\Delta y_{ij}, and bb terms in Eq. 6 only add per-row and per-column constants to SS, they aren't seen by the Sinkhorn normalization in training or the assignment at inference.
  2. Can you tune ArcGIS Pro parameters on the validation set and report the result?
  3. What do two trivial baselines score: (1) outputting the source buildings unchanged, and (2) the convex hull of the source buildings in each block?
  4. Can you retrain the raster baselines with the roads as a second input channel?
  5. Why do the no-road configurations output nothing for most single-building blocks (Sec 4.3)?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 9oc3

We thank Reviewer 9oc3 for the comments. Point-by-point replies below.

(a) / Q4. Raster baselines with a road channel. We retrained all three raster baselines with a road channel. Test results († new):

MethodPQPoly. IoUBIoUTFD↓Count acc.
CartoGenS (paper)0.6160.6320.6610.5090.759
RoofX-Net → + road†0.527 → 0.6680.570 → 0.6980.396 → 0.6240.631 → 0.5420.752 → 0.793
MSEONet → + road†0.492 → 0.6340.544 → 0.6700.373 → 0.5710.662 → 0.5670.727 → 0.774
SAM-BG → + road†0.380 → 0.4850.568 → 0.6760.382 → 0.5740.684 → 0.5930.524 → 0.627

Adding roads raises PQ by 0.10–0.14 for all three models, consistent with CartoGenS's own ablation (removing roads drops PQ from 0.512 to 0.080). This shows from another angle that, whatever the architecture, geographic context such as roads is necessary for building generalization, which is the core point of the paper. With roads, RoofX-Net is higher than CartoGenS in PQ, polygon IoU and count accuracy, and we have adjusted our conclusion against raster methods accordingly: road-aware RoofX-Net is higher on these, while CartoGenS has better boundaries and shape (better BIoU and TFD; spurious-hole rate 0.106 vs. 0.205), and outputs 1.31× the reference vertex count against about 31× for the raster outputs. Per category, CartoGenS has the higher PQ on single-building scenes (0.811 vs. 0.796 without holes), and RoofX-Net + road on multi-building scenes (0.536 vs. 0.424).

(b) / Q2. Tuned ArcGIS. We tuned ArcGIS Pro 3.7 (Aggregate Polygons + Simplify Building, roads as barriers) on the full validation split: first 32 settings over aggregation distance (5–40 m), simplification tolerance (2.5/5 m) and orthogonality, then 22 settings over minimum hole area (0–400 m²) and minimum building area (0–200 m²) at the two best aggregation distances. The setting with the best validation PQ was run once on test: PQ 0.493 (polygon IoU 0.482, BIoU 0.343), against 0.616 (0.632, 0.661) for CartoGenS.

Q3. Trivial baselines. Test results:

MethodPQPoly. IoUBIoU
Source buildings unchanged0.0650.3800.322
Convex hull per block0.4170.3970.262
Morphological closing0.3710.5080.303
CartoGenS (paper)0.6160.6320.661

Even on single-building scenes the target is not a cleaned-up union of the input: the convex hull gets PQ 0.673 on that category, CartoGenS 0.811.

Ablation / Q5. Why most outputs are empty without roads. Following QwQ3's suggestion, we ran a dummy-token control (same token count, all road coordinates set to zero). Test results:

BaseBase + dummyBase + roads
Empty60.9%48.7%19.4%
PQ0.0800.1160.512

Token count explains only a small part; most of the effect comes from road geometry. PolyWorld collapses in the same way without its road channel (95.1% empty, PQ 0.019): without roads its corner detector doesn't learn the target corners (precision 0.04–0.09, vs. 0.30 with roads). An empty output is a degenerate solution of the successor objective, where every query points to itself; the model falls into it when the scene encoding is not enough to pin down the target vertices and the matched rows can't be learned.

Q1. Is gijg_{ij} an edge-length penalty? Judging from the learned weights, it acts as a learnable edge-length prior, and this effect cannot be replaced by adding a hand-crafted distance penalty at inference. Δxij\Delta x_{ij}, Δyij\Delta y_{ij} and the bias cancel in the assignment, and the learned prior is −1.79∣Δx∣−2.14∣Δy∣−1.57d-1.79|\Delta x|-2.14|\Delta y|-1.57d. Adding it only at inference actually lowers test PQ from 0.561 to 0.456; the benefit comes from training it jointly with the scores (0.616 with gg in training vs. 0.561 without).

Reviewer_9oc3OpenReview ↗

Response to Authors

I thank the authors for their responses. My concerns regarding ArcGIS tuning, trivial non-learning baselines, and the role of the edge-length penalty have been resolved. My remaining concern is the comparison with the raster baselines under same conditioning modalities. Specifically, RoofX-Net leads on multiple-building scenes where scene-level formulation should be most important. I find the benchmark and problem formulation useful, but the current paper is framed around the model, and the central claim of it's performance no longer holds. Therefore, I will keep my score.

Reviewer_bpnGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a new approach for building outline extraction, which is able to handle holes inside a building. Source building and road vertices become tokens (predecessor/self/successor coordinates, a type flag, and a scalar scene-scale feature) and pass through a transformer encoder. A DETR-style decoder with 256 learned queries predicts a shared vertex pool. A PolyWorld-style successor matrix, trained with Sinkhorn and decoded by max-weight assignment, links vertices into cycles. Cycle orientation separates exterior from interior rings, so building count and hole structure come out of the prediction rather than being inherited from the source. A learned “relative-geometry” bias is added to the successor scores. To conduct experiments, authors also create a dataset from OS Open Zoomstack, and compare the proposed method with several recent approaches and frontier AI model (e.g. GPT-6 Astra). Overall, this paper is clearly written and the idea sounds interesting, however, I have concerns on the experiments.

Critical strengths
  • Clean and new formulation. Encoding all target rings as cycles of one permutation over a single shared vertex pool is elegant. Merges, splits, and holes need no per-building query groups (unlike RoomFormer) and no source-to-target correspondence. This is a new approach for building outline extraction.

  • Sound benchmark design. The geographically disjoint splits, the four-way categorization by target structure, and the reporting of instance, boundary, hole, and count metrics alongside area overlap are all good. The dataset could be a useful community resource. Better results are shown, compared with several other approaches.

  • Reproducibility. Appendices B–D give enough detail to reimplement the method and the metrics.

Critical weaknesses
  • Methodological novelty is limited for ICLR. The decoder is essentially DETR set prediction plus PolyWorld’s successor machinery: coordinates concatenated to descriptors, an attention GNN, ScoreNet, and Sinkhorn.

  • More comparison is needed, for example Pix2poly. Besides, why not show results on popular datasets such as WHU, INRIA datasets? INRIA also includes many buildings with holes. Although this paper presents a new dataset, the results on existing popular datasets are always helpful.

  • Missing reference to recent approach [a], which also directly generate building outline geometry. [a] Rethinking language models for building outline extraction from remote sensing imagery, CVPRW 2026.

  • What is PQ metric? How is this metric defined? In building outline extraction, there are many other popular metrics on the quality of outlines.

Questions for authors
  1. How were the GPT in-context examples selected, and how many were used?

  2. What is PQ metric? Any reference papers?

  3. Is the relative-geometry term effective only because it shapes training through Sinkhorn? What happens if it is applied only at inference, or replaced by a fixed −λ·d term?

1 comment
AuthorsOpenReview ↗

Response to Reviewer bpnG

We thank Reviewer bpnG for the comments. Point-by-point replies below.

On the task as described in the summary. We'd like to clarify this first: the paper does not extract building outlines from remote-sensing imagery. The input is a 1:10,000 vector map (building polygons and road centrelines), and the output is the 1:25,000 vector map of the same area, i.e. cartographic generalization. Holes, merges and splits appear because the 1:25,000 map generalizes the 1:10,000 map; they have nothing to do with imagery.

W1. Novelty (DETR + PolyWorld). The decoder components do come from DETR and PolyWorld, and the paper says so. Our contribution is mainly in how the task is formulated: building generalization as scene-level vector generation, where the number of polygons and their nesting are predicted from one shared vertex pool, with no source–target correspondence needed, conditioned on roads and the scene extent.

W2. Pix2Poly, WHU/INRIA. WHU and INRIA are image datasets with no source vector maps, so they can't be used for this task. As the closest polygon-prediction baseline, we trained PolyWorld on this task, following the multi-stage schedule of the original paper, with building and road rasters as input. Test results:

MethodPQPoly. IoUBIoUTFD↓Raw validitySelf-int.
CartoGenS (paper)0.6160.6320.6610.50970.4%10.3%
PolyWorld†0.5760.6280.5910.55369.9%21.4%

W3. Reference [a]. Thanks for pointing it out; it is related work on image-based polygon prediction.

W4 / Q2. Definition of PQ. PQ is panoptic quality (Kirillov et al., CVPR 2019), computed over building instances:

PQ=∑(p,g)∈TPIoU(p,g)∣TP∣+12∣FP∣+12∣FN∣,\mathrm{PQ}=\frac{\sum_{(p,g)\in TP}\mathrm{IoU}(p,g)}{|TP|+\frac{1}{2}|FP|+\frac{1}{2}|FN|},

with one-to-one matching at IoU>0.5\mathrm{IoU}>0.5; polygons include their holes. Besides PQ we report polygon IoU, boundary IoU, TFD (turning-function distance), hole IoU, spurious-hole rate, count accuracy and vertex-count ratio, all defined in the paper's appendix.

Q1. In-context examples for GPT. Each query gets two examples, retrieved from a fixed pool of 256 training scenes by nearest neighbour on seven log-scaled input descriptors (building count, vertex count, hole count, road segment count, total building area, total perimeter, coordinate scale), z-scored over the pool. To check whether this choice matters, we compared options on a stratified validation subset of 400 scenes (100 per category):

In-context examplesNoneRandom 2Retrieved 2 (paper)Retrieved 3
PQ0.0750.3840.4420.454

Retrieval beats random by 0.058 PQ (paired bootstrap 95% CI [0.024, 0.092]); a third example adds only 0.012 ([−0.007, 0.031]).

Q3. Relative geometry term. Judging from the learned weights, it acts as a learnable edge-length prior, and this effect cannot be replaced by adding a hand-crafted distance penalty at inference. Δxij\Delta x_{ij}, Δyij\Delta y_{ij} and the bias only add constants to each row and column, which cancel in the assignment; the learned prior is −1.79∣Δx∣−2.14∣Δy∣−1.57d-1.79|\Delta x|-2.14|\Delta y|-1.57d. Taking a model trained without this term and adding it only at inference lowers test PQ from 0.561 to 0.456 and raises empty outputs from 11.1% to 27.6%; replacing it at inference with a fixed −λd-\lambda d (λ\lambda chosen on validation) gives 0.555. So the benefit comes from training it jointly with the scores (0.616 with gg in training vs. 0.561 without), not from the geometric prior at inference itself.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (1/2): task clarification and new results

We thank the three reviewers for their comments. During the response period we tried to run the experiments you asked for. All CartoGenS numbers below are the ones in the paper. The new experiments use the same evaluation code, the same test split and a fixed seed; the settings of every new run were fixed on the validation split, and the test split was run once. Rows marked † are new.

What the task is. CartoGenS does not extract building outlines from imagery. The input is a 1:10,000 vector scene (building polygons and road centrelines from OS Open Zoomstack), and the output is the 1:25,000 vector scene of the same area as published by Ordnance Survey. This is cartographic generalization: buildings are merged, simplified, removed or kept, holes appear or disappear, and there is no given correspondence between output and input polygons. No imagery is involved, so image datasets such as WHU or INRIA do not fit this task.

Summary of new test results. B: buildings, R: roads. BIoU: boundary IoU (raster domain). TFD: turning-function distance (lower is better). VCR: predicted / reference vertex count. Valid: raw scene validity, no repair. SI: scenes with a self-intersecting ring. Empty: scenes with no output polygon.

MethodInputPQPoly. IoUBIoUTFD↓Count acc.VCRValidSI↓Empty↓
CartoGenS (paper)B+R vector0.6160.6320.6610.5090.7591.3170.4%10.3%3.9%
PolyWorld (trained on this task)†B+R raster0.5760.6280.5910.5530.7250.8569.9%21.4%2.9%
PolyWorld, no road†B raster0.0190.0160.0120.7910.0170.523.1%21.3%95.1%
RoofX-Net (paper)B raster0.5270.5700.3960.6310.75231.199.9%00
RoofX-Net + road†B+R raster0.6680.6980.6240.5420.79330.799.7%00
MSEONet (paper)B raster0.4920.5440.3730.6620.72733.099.8%00
MSEONet + road†B+R raster0.6340.6700.5710.5670.77431.299.8%00
SAM-BG (paper)B raster0.3800.5680.3820.6840.52433.889.3%00
SAM-BG + road†B+R raster0.4850.6760.5740.5930.62732.790.3%00
GPT-6-Astra (paper)B+R vector0.5870.6040.5660.5000.7710.8797.6%0.4%0.9%
ArcGIS Pro, paper settingB+R vector0.4980.5020.3530.6660.7861.44100%00
ArcGIS Pro, validation-tuned†B+R vector0.4930.4820.3430.6610.7911.31100%00
Identity (source unchanged)†–0.0650.3800.3220.7540.0300.89100%00
Convex hull per block†–0.4170.3970.2620.6220.7640.84100%00
Morphological closing†–0.3710.5080.3030.6990.6881.2099.6%00.3%

Against raster baselines with roads. With a road channel, RoofX-Net is higher than CartoGenS in PQ (0.668 vs. 0.616), polygon IoU (0.698 vs. 0.632), raster IoU (0.851 vs. 0.822), F1 (0.920 vs. 0.902) and count accuracy (0.793 vs. 0.759). With this comparison added, we have adjusted our conclusion against raster methods accordingly: road-aware RoofX-Net is higher in PQ and polygon IoU, and CartoGenS is stronger on boundaries and shape. This result also supports the paper's core point from another angle: building generalization needs geographic context such as roads. Adding roads raises PQ by 0.10–0.14 for all three raster models; in the paper's ablation, removing roads drops CartoGenS from 0.512 to 0.080; and PolyWorld without roads outputs almost nothing (PQ 0.019). Whatever the architecture, roads are decisive. On boundaries and shape: boundary IoU 0.661 vs. 0.624, SSIM 0.913 vs. 0.903, TFD 0.509 vs. 0.542, spurious-hole rate 0.106 vs. 0.205. It outputs 1.31× the reference vertex count, while the raster outputs have about 31×. Per category, CartoGenS has the higher PQ on single-building scenes (0.811 vs. 0.796 without holes, 0.790 vs. 0.779 with holes), and RoofX-Net + road on multi-building scenes (0.536 vs. 0.424, 0.526 vs. 0.367).

AuthorsOpenReview ↗

General response (2/2): topology of the raw output and next steps

Topology of the raw output. We added an analysis of the unrepaired output to locate where errors come from. For the paper model, 70.4% of raw test scenes are valid; scene by scene, 16.4% have a self-intersecting ring, 10.1% have overlapping exterior rings, 4.4% have a hole outside every exterior ring, and 3.7% are empty (a scene can fall into several of these). Self-intersections come mainly from two sources:

  1. The Hungarian matching used in training looks at coordinates only. When two queries converge on the same target vertex, which one gets the label is close to random, so the successor head gets contradictory supervision on whether to use that point, and at inference both queries can end up in the ring. On validation, among target vertices with no other target vertex nearby, 48% have two or more queries crowded around them, and rings containing such a duplicate self-intersect 5× as often as other rings.
  2. 46.6% of reference vertices are densified arc points: curves and rounded corners are stored as polylines with 0.3–0.4 m segments, which is 0.015 mm on a 1:25,000 map. This spacing is below the model's coordinate error, so the order of neighbouring arc points is flipped on 19.4% of arc edges, against 2.4% of other edges. Per 100 m of boundary, crossings on arcs are 10× as frequent as on straight walls (0.32 vs. 0.031).

Errors are concentrated in complex scenes: 60% of scenes with 4 or more target buildings have a self-intersecting ring, against 10% of single-building scenes.

Why other methods have fewer topology errors. Raster methods are per-pixel classifiers; their polygons are traced from the mask in post-processing and cannot self-intersect. Also, rule-based, LLM and raster methods mostly do not draw these densified arcs, and arcs are exactly where the crossings happen. The table shows how often each method reproduces target arc points on the test split (an arc point counts as reproduced if an output vertex lies within 0.5 m):

MethodTarget arc points reproducedOutput vertices on target arcs
CartoGenS (paper)65.7%28.6%
RoofX-Net + road18.0%*0.3%
GPT-6-Astra12.6%4.4%
PolyWorld6.8%2.1%
ArcGIS Pro, validation-tuned0.5%0.1%
Convex hull0.1%0.1%

*Pixel-traced polygons with about 31× the reference vertex count; the hits come from staircase vertices.

CartoGenS reproduces two thirds of the arc points, and that is also where its crossings are. Other methods don't fail on arcs because they don't try to represent them. Pixel metrics also start from a high floor: just copying the input gives raster F1 0.667 and SSIM 0.793 (the best raster model reaches 0.920 and 0.903). PolyWorld, which like CartoGenS predicts vertices and their connections, has 69.9% validity and 21.4% self-intersections (CartoGenS: 70.4% and 10.3%).

Improvements addressing these issues. We have improved the method to address the causes identified above, and will include the comparisons above when we resubmit.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.