CityGen Urban Building Image Dataset

This dataset is from the paperCityDreamer: Compositional Generative Model of Unbounded 3D Cities"The datasets constructed include OSM dataset and Google Earth dataset, and the results have been included in CVPR 2024.

1. OSM dataset

The dataset comes from OpenStreetMap (OSM). It includes rasterized semantic maps and height fields of 80 cities, covering an area of more than 6,000 square kilometers. During the rasterization process, the vector geometry information was converted into images, using the EPSG:3857 coordinate system and zoomed in to level 18, which is about 0.597 meters per pixel.

The semantic map uses different colors to represent elements in different locations, such as red for roads, yellow for buildings, green for green land, cyan for construction sites, and blue for water areas. The height field mainly represents the height of buildings, with the road height set to 4, the water height set to 0, and the tree height randomly sampled between 8 and 16.

2. Google Earth Dataset

The dataset is collected from Google Earth Studio, specifically the Manhattan and Brooklyn areas of New York City. It contains 400 tracks, each track contains 60 images, and a total of 24,000 real city images. The radius of the tracks ranges from 125 meters to 813 meters, and the height ranges from 112 meters to 884 meters. Google Earth Studio provides the intrinsic and extrinsic parameters of the camera, allowing automatic annotations to be generated for semantic and building instance segmentation. Individual building instances are identified by detecting connected components in the semantic graph, and annotations are generated by projecting the city layout onto the image using the camera parameters.

Multi-view consistency: This dataset includes real city images from different viewpoints, which enhances the ability of neural networks to generate cities from different viewpoints.