Figure 1: Talinn, the largest city in the Building3D dataset, along with building and house point clouds, mesh and wireframe models.
To facilitate the development of real-world 3D reconstruction in the field of surveying and mapping, we propose Building3D (Building3D: An Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds), a high-quality aerial point cloud-based house modeling dataset. This dataset has three main advantages:
-
Real World: Different from existing artificially constructed 3D building datasets, it consists of real-world buildings in the Republic of Estonia.
-
Rich categories: about 160,000 3D point cloud building data covering more than 100 housing types.
-
Rich annotations: including surface point clouds of buildings and roofs, mesh and wireframe models.
Building3D is currently the largest real-world 3D model dataset in the academic world, providing a broad space for future research on real-world 3D white model modeling. Using this data set, we explored the robustness and generalization of various academic tasks such as house surface reconstruction, point cloud completion, denoising and outlier removal, and 3D model generation, and made many valuable discoveries and verified Its development and application prospects from noise removal and outlier removal, 3D reconstruction to 3D model generation are discussed. We hope that building3D and the corresponding algorithmic benchmarks will bring new challenges and opportunities for academic research and industrial applications.
-
Project page: http://building3d.ucalgary.ca/
-
Paper: https://arxiv.org/abs/2307.11914
-
Dataset Download:
-
Entry-level datasets are now available.
-
Tallinn city data will be released before 8.15
-
The test server will also be available until 8.15
-
3D reconstruction and model generation for the real world are issues that have been receiving much attention in the field of surveying and mapping and smart city construction, and have made some progress in recent years. However, due to the lack of large-scale real-world 3D building databases in both academia and industry, most technical approaches still rely on simulation and the limited-quality ROOFN3D [1] dataset. Therefore, the community urgently needs a large-scale and high-quality real-world 3D building dataset, which will help advance many 3D reconstruction tasks and downstream applications.
Dataset Features
Building3D provides three modal information for each building, including: building and roof point cloud, building and roof mesh model and building and roof wireframe model. It has more than 100 house models. Figure 2 below shows some classic house types and their corresponding point clouds, wireframes (top view and side view), roof and wireframe mesh models.
Figure 2: The point cloud corresponding to the typical 10 house types in the Building3D dataset, the wireframe of the top view and the side view, and the mesh model corresponding to the house and the wireframe. whaosoft aiot http://143ai.com
At the same time, due to the data set in the real world, there may also be cases where the point cloud is sparse, the corner points of the building disappear, and the number of corner points of the building is too large.
downstream application
Building3D has brought a wide range of exploration space to the academic community. In this paper, we selected the 3D building reconstruction task shown in Figure 3 for evaluation and analysis. We propose a unified supervised and self-supervised end-to-end house modeling framework, which divides the 3D reconstruction task into two parts: 1) building edge detection and recognition; 2) building effective edge connection work.
Figure 3: 3D building reconstruction results
future work
Regarding the dataset itself, we are committed to continuously expanding and updating the dataset to meet broader research needs. In addition to existing applications, we also plan to further develop other downstream tasks, such as sparse point cloud completion, semantic segmentation and how to generate corner points when corner points disappear, and improve the accuracy of corner point detection.