BlockPlanner: City Block Generation with
Vectorized Graph Representation ICCV 2021

Linning Xu*1     Yuanbo Xiangli*1     Anyi Rao1     Nanxuan Zhao1,3     Bo Dai2     Ziwei Liu2     Dahua Lin1,3,4
*denotes equal contribution



City modeling is the foundation for computational urban planning, navigation, and entertainment. In this work, we present the first generative model of city blocks named BlockPlanner, and showcase its ability to synthesize valid city blocks with varying land lots configurations. We propose a novel vectorized city block representation utilizing a ring topology and a two-tier graph to capture the global and local structures of a city block. Each land lot is abstracted into a vector representation covering both its 3D geometry and land use semantics. Such vectorized representation enables us to deploy a lightweight network to capture the underlying distribution of land lots configuration in a city block. To enforce intrinsic spatial constraints of a valid city block, a set of effective loss functions are imposed to shape rational results. We contribute a pilot city block dataset to demonstrate the effectiveness and efficiency of our representation and framework over the state-of-the-art. Notably, our BlockPlanner is also able to edit and manipulate city blocks, enabling several useful applications, e.g., topology refinement and footprint generation.



Overview of BlockPlanner. The model learns to generate city blocks under a VAE training scheme. Each city block is represented by a graph with lots arranged in a ring topology. After encoding the input graph into a 128-d latent code, the decoder first predicts 1) the aspect ratio of the block shape, 2) the edge map indicating adjacent relations, and 3) the initial feature for each lot node. Then the initial features are updated through iterative message passing and output 1) existence of lots, 2) the land use category, 3) the geometry parameters, and 4) the merged and boundary attributes, with four linear heads. These attributes are further used to calculate the reconstruction error and the geometry violations, plus the conventional variational loss.

Example Results


Qualitative evaluations. Column 2-3: LayoutVAE and MolVAE fail to capture valid block topology, while our BlockPlanner efficiently generates diverse and valid block layouts with reasonable land use categories. Column 5-6: marker (L) indicates land use level view. Though pixel-based BlockGAN generates sensible structure for large lots, the artifacts such as the fuzzy boundary are significant. In contrast, our vectorized representation captures much more accurate geometry even in small areas. (Dimgray color indicates the unfilled regions.)



The website template was borrowed from Michaƫl Gharbi.