ACPV-Net: All-Class Polygonal Vectorization for Seamless Vector Map Generation from Aerial Imagery

Weiqin Jiao; Hao Cheng; George Vosselman; Claudio Persello

cs.CVMar 17, 2026

ACPV-Net: All-Class Polygonal Vectorization for Seamless Vector Map Generation from Aerial Imagery

Authors: Weiqin Jiao, Hao Cheng, George Vosselman, Claudio Persello

Organizations: Faculty of Geo-Information Science and Earth Observation (ITC) University of Twente, The Netherlands

Abstract

We tackle the problem of generating a complete vector map representation from aerial imagery in a single run: producing polygons for all land-cover classes with shared boundaries and without gaps or overlaps. Existing polygonization methods are typically class-specific; extending them to multiple classes via per-class runs commonly leads to topological inconsistencies, such as duplicated edges, gaps, and overlaps. We formalize this new task as All-Class Polygonal Vectorization (ACPV) and release the first public benchmark, Deventer-512, with standardized metrics jointly evaluating semantic fidelity, geometric accuracy, vertex efficiency, per-class topological fidelity and global topological consistency. To realize ACPV, we propose ACPV-Net, a unified framework introducing a novel Semantically Supervised Conditioning (SSC) mechanism coupling semantic perception with geometric primitive generation, along with a topological reconstruction that enforces shared-edge consistency by design. While enforcing such strict topological constraints, ACPV-Net surpasses all class-specific baselines in polygon quality across classes on Deventer-512. It also applies to single-class polygonal vectorization without any architectural modification, achieving the best-reported results on WHU-Building. Data, code, and models will be released at: https://github.com/HeinzJiao/ACPV-Net.

Figures & tables

Figure 2: Overview of ACPV-Net. It unifies semantically supervised conditioning and proposition-driven topological reconstruction: the former produces coherent semantic–geometric evidence through diffusion-based vertex generation under semantic supervision, the latter deterministically reconstructs a topology-consistent vector basemap via overdense PSLG construction and vertex-guided subset selection.

Dataset	Multi-class	V	S	T	C
Inria [ 24 ]	✗	✗	✗	✗	✓
WHU-Building [ 14 ]	✗	✓	✗	✗	✓
ISPRS Vaihingen [ 13 ]	✓	✗	✗	✗	✗
ISPRS Potsdam [ 12 ]	✓	✗	✗	✗	✗
LoveDA [ 33 ]	✓	✗	✗	✗	✗
Deventer-512 (Ours)	✓	✓	✓	✓	✓

Table 1: Comparison of representative datasets for polygonal outline extraction (top) and land-cover mapping (bottom). “V” indicates vector annotations; “S” shared-edge consistency; “T” global topology-consistency metrics; and “C” cadastral-aligned boundaries (as opposed to purely visual contours).

Method	Gap $\downarrow$	Inter $\downarrow$	Intra $\downarrow$	Shared $\uparrow$
DeepSnake [ 29 ] (CVPR ’20)	12.41	68.86	51.16	38.73
FFL [ 7 ] (CVPR ’21)	5.48	29.17	0.08	9.20
TopDiG [ 38 ] (CVPR ’23)	8.47	13.57	0.00	10.81
HiSup [ 36 ] (ISPRS ’23)	5.43	4.50	0.00	25.73
GCP [ 40 ] (TGRS ’25)	8.75	10.39	41.91	20.25
ACPV-Net (Ours)	0.00	0.00	0.00	100.00

Table 2: Global topological consistency on Deventer-512. Metrics include gap rate (Gap), inter-class overlap rate (Inter), intra-class overlap rate (Intra) and shared-edge consistency rate (Shared). Best and second-best values are highlighted in bold and underlined , respectively. All percentage values are reported to two decimal places.

Class	Method	Semantic fidelity		Vertex efficiency		Geometric accuracy		Topological fidelity
Class	Method	IoU $\uparrow$	B-IoU $\uparrow$	N-ratio $\rightarrow$ 1	C-IoU $\uparrow$	PoLiS $\downarrow$	MTA $\downarrow$	APLS $\uparrow$	$\chi$ -Err $\downarrow$	$\beta$ -Err $\downarrow$
Road	DeepSnake [ 29 ] (CVPR ’20)	10.35	7.90	74.91	4.58	33.18	53.38	64.85	7.82	10.11
	FFL [ 7 ] (CVPR ’21)	62.00	58.08	3.74	42.28	11.48	44.05	78.53	15.31	15.40
	TopDiG [ 38 ] (CVPR ’23)	58.75	55.95	1.80	44.06	10.28	44.27	70.65	5.90	6.54
	HiSup [ 36 ] (ISPRS ’23)	73.92	69.29	2.00	57.36	4.97	44.84	81.06	5.16	5.45
	GCP [ 40 ] (TGRS ’25)	59.93	56.58	0.86	44.96	9.18	49.19	73.00	4.38	5.18

Table 3: Quantitative comparison on Deventer-512. Best and second-best values are highlighted in bold and underlined , respectively. APLS is evaluated only on elongated classes (roads and water bodies).

Panel of Figure 3 — Figure 3: Qualitative comparison on Deventer-512. The three rows show representative urban, suburban, and rural scenes, respectively. From left to right: aerial imagery, ground truth, TopDiG [ 38 ] , HiSup [ 36 ] , and Ours. Land-cover classes are color-coded; polygon outlines are drawn in black, vertices are highlighted with orange dots, and inter-class overlaps and gaps are marked in red and black, respectively.

Method	Sem. fidelity		Vertex efficiency		Geo. accuracy		Topo. fidelity
Method	IoU $\uparrow$	B-IoU $\uparrow$	N-ratio $\rightarrow$ 1	C-IoU $\uparrow$	PoLiS $\downarrow$	MTA $\downarrow$	$\chi$ -Err $\downarrow$	$\beta$ -Err $\downarrow$
DeepSnake [ 29 ]	71.39	64.10	2.05	54.21	2.17	46.85	4.43	6.29
FFL [ 7 ]	80.86	74.63	3.68	46.09	2.13	38.86	12.33	14.88
TopDiG [ 38 ]	83.79	77.73	1.95	62.25	1.99	38.82	2.13	2.56
HiSup [ 36 ]	87.63	82.82	1.93	67.15	1.40	35.27	1.95	2.28
GCP [ 40 ]	83.28	77.58	2.11	60.51	1.66	34.96	3.20	4.39

Table 4: Quantitative comparison on WHU-Building. Best values are in bold and second-best values are underlined .

Figure 4: Vertex activations under weak/ambiguous visual cues and along smooth boundaries (cartographic convention cases). From left to right: aerial image, pure discriminative decoding, without semantic supervision (No-SSC), ours, and ground truth.

Variant	V2B@2 $\uparrow$	V2B@4 $\uparrow$	V2B@6 $\uparrow$	Avg@2,4,6 $\uparrow$
No-SSC	0.38	0.48	0.53	0.46
Full SSC (ours)	0.78	0.87	0.90	0.85

Table 5: Vertex–Boundary Alignment (V2B@

\delta

) comparison between No-SSC and Full SSC . Global boundaries are computed as multi-class label transitions from the raster mask. Avg(2,4,6) denotes the mean of V2B@

2

, @

4

, and @

6

Panel of Figure 5 — Figure 5: Vertex peak-shape comparison between the pure discriminative baseline and our distributional reconstruction . Lower values of FWHM and [email protected] indicate sharper and more compact peaks, while higher Sharpness reflects stronger local contrast. The blue/red bar denotes the median/90th percentile.

Metric	Pure discrim.		Ours (dist. model)
	Median	$P_{90}$	Median	$P_{90}$
FWHM x $\downarrow$	9.55	10.23	6.09	7.64
FWHM y $\downarrow$	9.46	10.25	6.09	7.70
[email protected] $\downarrow$	75.0	105.0	21.0	39.0
Sharpness $\uparrow$	0.10	0.14	0.24	0.27

Table 6: Summary statistics (median and 90th percentile) of vertex peak-shape metrics.

Method	Gap $\downarrow$	Inter $\downarrow$	Intra $\downarrow$	Shared $\uparrow$
No PSLG + DP (w/o VSS)	1.04	0.14	0.00	15.38
PSLG + DP (w/o VSS)	0.00	0.00	0.00	100.00

Table 7: Ablation of Topology-consistency on the reconstruction stage. Metrics are averaged over DP tolerances

\varepsilon\!\in\!\{1,2,3,4\}

PolyTopoBench: A Benchmark for Complex Vector Polygon Generation from Remote Sensing Imagery
Sep 26, 2026Zeping Liu, Ni Lao, Weiwei Sun +5Remote Sensing Benchmarks
The University of Texas at Austin · Amazon · University of Maryland +1
Vector Map as Language: Toward Unified Remote Sensing Vector Mapping
Jun 9, 2026Yinglong Yan, Yunkai Yang, Haoyi Wang +7Remote Sensing Image Understanding Vision-Language Models
school of Artificial Intelligence and Robotics, Hunan University, Changsha 410082, China · Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong · Department of Geomatics Engineering, Changsha University of Science and Technology, Changsha 410114, China +1
Topologically Consistent Agricultural Parcel Vectorization with Semantic-Guided Diffusion and Topology-Aware Polygonization
Sep 7, 2026Weiqin Jiao, Xiaolong Zuo, Claudio PerselloPrecision Agriculture Diffusion Models
Department of Earth Observation Science, Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, Enschede, 7522 NH, The Netherlands · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, 430072, China

Sep 26, 2026J/K move · Enter open · S save

PolyTopoBench: A Benchmark for Complex Vector Polygon Generation from Remote Sensing Imagery

Zeping Liu, Ni Lao, Weiwei Sun, Gil Wolff+4

The University of Texas at Austin · Amazon · University of Maryland+1

Remote Sensing Benchmarks

Open paper AskPDF arXiv

Vector polygon generation converts visual inputs, e.g., remote sensing (RS) images, into vectorized polygonal geometries, supporting applications such as autonomous driving, vector map construction, and remote sensing. Early pipelines predict raster masks and post-process them into polygons, which prevents end-to-end optimization and may miss small objects or introduce inaccurate vertices. Recent methods directly generate vector polygons, but most focus on simple exterior contours, while they either cannot represent complex polygons with holes or fail to preserve their topology. In this paper, we propose PolyTopoBench, a unified evaluation framework for vector polygon generation from RS images with explicit emphasis on complex polygons. PolyTopoBench evaluates both exterior and interior rings, and benchmarks 11 representative methods, including segmentation-based polygonization pipelines, vision foundation model baselines, and specialized vector polygon generators, on two RS-image datasets covering buildings, roads, vegetation, and unvegetated regions. Experiments show that existing methods often recover simple exterior boundaries but degrade substantially on polygons with holes or multiple rings. These results reveal complex polygon generation as an unresolved challenge and motivate topology-aware benchmarks and model designs. Code and data are available at https://github.com/seai-lab/PolyTopoBench.

Abstract

Explore similar work

PolyTopoBench: A Benchmark for Complex Vector Polygon Generation from Remote Sensing Imagery

Vector Map as Language: Toward Unified Remote Sensing Vector Mapping

Topologically Consistent Agricultural Parcel Vectorization with Semantic-Guided Diffusion and Topology-Aware Polygonization