RealCity3D: A Large-scale Georeferenced
3D Shape Dataset of Real-world Cities

1 New York University, 2 Tsinghua University
* Equal contributions.
Equal contributions.
The corresponding author is Chen Feng


Existing 3D shape datasets in the research community are generally limited to objects or scenes at the home level. City level shape datasets are rare due to the difficulty in data collection and processing. However, such datasets uniquely present a new type of 3D data with a high variance in geometric complexity and spatial layout styles, such as residential/historical/commercial buildings and skyscrapers. This work focuses on collecting such data, and proposes city generation as new tasks for data-driven content generation. Thus, we collect over 1,000,000 geo-referenced 3D building models from New York City and Zurich. We benchmark various baseline performance on two challenging tasks: (1) city layout generation, and (2) building shape generation. Moreover, we propose an auto-encoding tree neural network for 2D building footprint and 3D building cuboid generation. The dataset, tools, and algorithms will be released to the community.


The current RealCity3D contains more than 1,000,000 geo-referenced objects of New York City and Zurich. The four different representations are provided for each object, including polygon meshes, triangle meshes, point clouds, and voxel grids.

City Country Continent Polygon Meshes Voxel Triangulated Meshes Point Cloud
New York City United States North America download download download download
Zurich Switzerland Europe download download download download


Task 1: City layout generation.

Quantitative comparisons of city layout generation performance with various data-drivenbaseline methods.

City layout generation results of the models trained on NYC dataset.

3D generation results of AETree trained on the NYC Dataset.

Task 2: Building shape generation.

Benchmark of building point cloud generation on two datasets

Some testing results of FoldingNet trained on RealCity3D.

Some shape generation results of Latent-GAN trained on RealCity3D.


The authors gratefully acknowledge the useful comments and suggestions from Yuqiong Li and Tanay Varshney.