Reconstructing building wireframe from airborne LiDAR point clouds yields a compact, topology-centric representation that enables structural understanding beyond dense meshes. Yet a key limitation persists: conventional methods have failed to achieve accurate wireframe reconstruction in regions afflicted by significant noise, sparsity, or internal corners. This failure stems from the inability to establish an adaptive search space to effectively leverage the rich 3D geometry of large, sparse building point clouds. In this work, we address this challenge with Delaunay Canopy, which utilizes the Delaunay graph as a geometric prior to define a geometrically adaptive search space. Central to our approach is Delaunay Graph Scoring, which not only reconstructs the underlying geometric manifold but also yields region-wise curvature signatures to robustly guide the reconstruction. Built on this foundation, our corner and wire selection modules leverage the Delaunay-induced prior to focus on highly probable elements, thereby shaping the search space and enabling accurate prediction even in previously intractable regions. Extensive experiments on the Building3D Tallinn city and entry-level datasets demonstrate state-of-the-art wireframe reconstruction, delivering accurate predictions across diverse and complex building geometries.
Figures & tables
Figure 2: Comparison of geometric context between 2D heightmap and Delaunay scoring. While the 2D heightmap collapses 3D structural details and obscures internal corners (red box), our Delaunay scoring effectively preserves the underlying geometric context.
Figure 3: Visualization of the edge-wise dihedral angle and vertex-wise corner score , derived through Delaunay graph scoring across input point clouds, projected onto the Bird’s-Eye View (BEV). The results demonstrate a robust capability to accurately capture the underlying surface and the curvature of the corresponding region. Notably, the scoring adeptly represents even the most minute degrees of surface curvature (fourth column) and maintains efficacy even when the point cloud is highly sparse (fifth column).
Figure 4: Overall pipeline of Delaunay Canopy. The pipeline begins by utilizing Delaunay triangulation to construct the point cloud into a Delaunay graph. The method then sequentially computes metrics that represent the local curvature around each element, proceeding in the order of face, edge, and vertex. This process encapsulates rich 3D information and introduces a judicious search space into the pipeline.
Figure 5: Comparison between Corner Score Sampling (ours) and Farthest Point Sampling (FPS). Resultant point sets after sequentially sampling 500 points ( K = 500) and then 150 points ( K = 150) with each method are shown. For enhanced clarity, the ground truth wireframe is visualized. While FPS performs a uniform sampling, our corner score sampling method retains points located in areas of pronounced curvature ( e.g . , roof corners and edges), making them highly probable corner candidates.
Method
Tallinn City
Entry-Level
Distance
Corner
Wire
Distance
Corner
Wire
WED ↓
ACO ↓
CP ↑
CR ↑
CF1 ↑
EP ↑
ER ↑
EF1 ↑
WED ↓
ACO ↓
CP ↑
CR ↑
CF1 ↑
EP ↑
ER ↑
EF1 ↑
Building3D (PointNet) [ 27 ]
0.264
0.241
88.2
77.6
82.6
80.4
71.9
75.9
0.262
0.239
89.1
77.8
83.1
80.7
72.7
76.5
PC2WF [ 17 ]
0.554
0.508
21.4
50.5
30.1
2.8
16.7
4.8
0.490
0.422
28.5
16.3
20.7
3.1
19.1
5.3
Point2Roof [ 14 ]
0.260
0.236
89.3
78.5
83.6
81.2
72.2
76.4
0.257
0.238
89.5
78.1
83.4
81.4
73.4
77.2
PointTransformer ∗ [ 44 ]
0.257
0.238
89.7
79.1
84.1
81.3
72.4
76.6
0.259
0.233
89.7
79.4
84.2
81.5
73.8
77.5
Table 1: Quantitative comparisons on the Building3D Tallinn city and entry-level datasets. The notation MODEL∗ signifies a model acting as the feature extractor within the Point2Roof [ 14 ] pipeline. Best , second best , and third best are highlighted.
Figure 6: Qualitative comparison with other methods. The notation MODEL∗ signifies a model acting as the feature extractor within the Point2Roof [ 14 ] pipeline. The blue boxes illustrate the robustness of our framework, demonstrating its capability to reconstruct the wireframe even in regions that are severely sparse (first row), highly noisy (second row), or exhibit minute degrees of curvature (third row).
Figure 7: Qualitative comparison with the strongest baseline BWFormer [ 19 ] on challenging cases with complicated roof structures. The orange boxes show our method’s ability to fully leverage 3D geometric context, a capability often absent when only 2D information is employed, thus facilitating superior detection of interior corners in addition to those on the periphery.
Table 8Table 9
Exp.
Wire Selection
Prior-based
Distance
Corner
Wire
#
Module
Query Scaling
WED ↓
ACO ↓
CP ↑
CR ↑
F1 ↑
EP ↑
ER ↑
F1 ↑
1
Point2Roof
-
0.248
0.225
92.4
80.5
86.0
85.8
72.8
78.8
2
Ours
-
0.239
0.210
93.1
81.1
86.7
85.6
73.7
79.2
3
Ours
✓
0.232
0.194
94.5
82.8
88.3
87.1
74.0
80.0
Table 5: Analysis of geometric prior vs. architecture.
Figure 9: The resulting mesh, reconstructed from the wireframe output generated by the Delaunay Canopy. The wireframe demonstrably encapsulates sufficient geometric information, allowing the resultant mesh to exhibit a highly faithful form.
Method
Distance
Corner
Wire
WED ↓
ACO ↓
CP ↑
CR ↑
CF1 ↑
EP ↑
ER ↑
EF1 ↑
PointMAE † [ 26 ]
-
0.330
75.0
47.0
58.0
52.0
12.0
20.0
PointM2AE † [ 43 ]
-
0.320
79.0
58.0
67.0
50.0
7.0
12.0
Point2Roof [ 14 ]
-
0.390
65.0
30.0
41.0
66.0
8.0
14.0
Linear self-supervised [ 34 ]
-
0.350
70.0
60.0
65.0
67.0
16.0
25.0
Supervised [ 34 ]
-
0.290
90.0
53.0
66.0
88.0
23.0
36.0
Table 6: Performance evaluation based on the official Building3D [ 34 ] Tallinn city dataset leaderboard. Baseline metrics are sourced from the public leaderboard. The notation MODEL† designates models acting exclusively as the feature extraction backbone within the standard Building3D [ 34 ] architecture. Best , second best , and third best are highlighted.
Method
Distance
Corner
Edge
WED ↓
ACO ↓
CP ↑
CR ↑
CF1 ↑
EP ↑
ER ↑
EF1 ↑
Building3D (PointNet) [ 27 ]
0.271
0.245
87.8
77.9
82.5
80.8
71.5
75.9
PC2WF [ 17 ]
0.548
0.512
21.9
50.1
30.4
3.1
16.5
5.3
Point2Roof [ 14 ]
0.265
0.239
89.0
78.2
83.3
81.5
71.8
76.3
PointTransformer ∗ [ 44 ]
0.260
0.242
89.4
78.8
83.8
81.7
72.0
76.5
PointMLP ∗ [ 21 ]
0.258
0.236
90.3
79.3
84.4
82.1
72.5
77.0
Table 7: Robustness validation via 5-Fold Cross-Validation. The notation MODEL∗ signifies a model acting as the feature extractor within the Point2Roof [ 14 ] pipeline. Best and second best are highlighted.
Method
Distance
Corner
Edge
WED ↓
ACO ↓
CP ↑
CR ↑
CF1 ↑
EP ↑
ER ↑
EF1 ↑
Building3D (PointNet) [ 27 ]
0.407
0.354
78.1
68.9
73.2
72.6
64.2
68.2
PC2WF [ 17 ]
0.658
0.612
14.2
38.7
20.8
1.6
10.4
2.7
Point2Roof [ 14 ]
0.384
0.341
79.5
70.3
74.6
73.9
65.4
69.4
PointTransformer ∗ [ 44 ]
0.367
0.325
80.8
71.5
75.9
75.2
66.5
70.6
PointMLP ∗ [ 21 ]
0.359
0.318
81.3
72.0
76.4
75.6
67.1
71.1
Table 8: Robustness validation assessed against a perturbed dataset characterized by augmented sparsity and noise. The notation MODEL∗ signifies a model acting as the feature extractor within the Point2Roof [ 14 ] pipeline.
Figure 10: Visualization of performance differences between challenging (perturbed point cloud) and original inputs. The number next to Δ indicates the absolute performance difference, and the percentage (%) shows the relative decrease compared to the original performance. The notation MODEL∗ signifies a model acting as the feature extractor within the Point2Roof [ 14 ] pipeline.
Figure 11: Visualization of failure cases. Precise wireframe reconstruction is demonstrably compromised under several key scenarios. (a) When critical regions containing valid corner and wire features are unscanned and consequently absent from the point cloud. (b & c) When the intrinsic surface topology of the point cloud exhibits excessive smoothness. (d) When processing objects characterized by a closed topology, such as non-building volumetric structures.
Figure 12: Qualitative comparison with the strongest baseline BWFormer [ 19 ] . The orange bounding boxes specifically denote the regions where our method achieves a demonstrably superior reconstruction result.
Figure 13: Wireframe reconstruction results of Delaunay Canopy on Building3D dataset.
Figure 14: Wireframe reconstruction results of Delaunay Canopy on Building3D dataset.
We introduce BuildAnyPoint, a novel generative framework for structured 3D building reconstruction from point clouds with diverse distributions, such as those captured by airborne LiDAR and Structure-from-Motion. To recover artist-created building abstraction in this highly underconstrained setting, we capitalize on the role of explicit 3D generative priors in autoregressive mesh generation. Specifically, we design a Loosely Cascaded Diffusion Transformer (Loca-DiT) that initially recovers the underlying distribution from noisy or sparse points, followed by autoregressively encapsulating them into compact meshes. We first formulate distribution recovery as a conditional generation task by training latent diffusion models conditioned on input point clouds, and then tailor a decoder-only transformer for conditional autoregressive mesh generation based on the recovered point clouds. Our method delivers substantial qualitative and quantitative improvements over prior building abstraction methods. Furthermore, the effectiveness of our approach is evidenced by the strong performance of its recovered point clouds on building point cloud completion benchmarks, which exhibit improved surface accuracy and distribution uniformity.
We present WireframeDETR, our submission to the Structured Semantic 3D Reconstruction (S23DR) 2026 Challenge, which requires predicting a 3D building wireframe from multi-view COLMAP point clouds. Our method applies DETR-style set prediction directly to 3D point clouds, producing wireframes as sets of edge coordinate pairs without any intermediate vertex detection stage. We introduce three technical contributions: (1) contrastive denoising training that stabilises noisy Hungarian matching in early epochs; (2) a multi-scale encoder that aggregates the last encoder layer outputs via learned scalar weights; and (3) progressive auxiliary loss weighting that concentrates gradient signal on the decoder layers that most benefit from it. Our model achieves a public test HSS of 0.575 (F1~=0.664, IoU=~0.516) and a best validation HSS of 0.534 on the cleaned val split.
Generating compact polygonal models from point clouds is a key problem in 3D vision and computer graphics. However, due to inherent limitations of LiDAR scanning (e.g. range constraints and occlusions), critical scene information is often missing, leading to degraded reconstruction accuracy. To address this, we propose a plane assembling strategy that effectively recovers missing details while maintaining model compactness. We classify all the planes extracted from the scene into three categories: highly visible, barely visible, and invisible. The invisible planes, which are recovered by scene structure analysis, indicate the missing details. The three types of planes correspond to the three growth priorities. Each plane grows according to the priority level, and the space is partitioned progressively, namely, the hierarchical partition. Subsequently, we generate a watertight polygonal mesh from the partition via a min-cut-based optimization. Finally, comparisons on public datasets show the effectiveness and superiority of our method against mainstream approaches. The project page is available at https://hsr-3dv.github.io/.
Minjie Tang, Xiangfei Li
Independent Researcher · Huazhong University of Science and Technology