AETree: Areal Spatial Data Generation

Congcong Wen, Wenyu Han, Hang Zhao, Chen Feng

Keywords: content generation, spatial data representation, tree-based network

Abstract: Areal spatial data represent not only geographical locations but also sizes and shapes of physical objects such as buildings in a city. Data-driven generation of such vector-format data requires an effective representation. Inspired by the hierarchical nature of such spatial data, we propose AETree, a tree-based deep auto-encoder network. Unlike common strategies that either treat the data as an unordered set or sort them into a sequence, we preprocess the data into a binary tree via hierarchical clustering. Then a tree encoder learns to extract and merge spatial information from bottom-up iteratively. The resulting global representation is reversely decoded for reconstruction or generation. Experiments on large scale 2D/3D building datasets of both New York and Zurich showed superior performance of AETree than either set-based or sequential auto-regressive deep models.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

One-sentence Summary: Spatial hierarchy is a better representation than set or sequence for learning to generate spatial data.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=Vfq431xUi3