Latent Street Walk

Generative Urban Layouts To Guide Future Street Planning

2022

Keywords

Generative Urban Networks
Deep Learning
Variational Autoencoder (VAE)
‍

Tools

Google Colab
TensorFlow
OpenStreetMap

Advisor

Prof. Thomas Wortmann

Teammates

Joyce Wu

My Role

Design Computation-VAE Implementation
Concept Development

Highlights

Urban Network Generation: Leverages deep learning to generate urban from a 592-image dataset of the historical city layouts, offering a novel approach for future urban planning

Generative Approach: Applies Variational Autoencoders (VAE) to reconstruct existing street patterns and create new urban layouts.

Concept

How can we learn from the existing street networks and provide guidance for the future planning of the undeveloped or less developed areas?
‍

Urban street networks hold essential insights into an area's spatial patterns, profoundly influencing connectivity, accessibility, and infrastructure. Historically, diverse urban layout styles—such as medieval, grid, radial, and post-WW2—emerged to meet varying local needs for movement and interaction.
This project explores how deep learning models can analyze these existing layouts to generate new urban network maps.
‍

Workflow & Dataset

The research analyzed 65 European cities, each representing one of the four distinct urban layout styles: medieval, grid, radial, and post-WW2.

Using OpenStreetMap and Overpass Turbo (overpass-turbo.eu), 592 road network images were extracted to train a Variational Autoencoder (VAE) model.

Implemented in Colab with TensorFlow, the model aimed to reconstruct and generate urban street layouts by capturing latent patterns in the data.

VAE Implementation

The VAE was employed to generate new street network styles based on the dataset of urban layouts. The model captures high-level urban network metrics by encoding patterns into low-dimensional latent vectors.

The autoencoder is trained to:
‍
1) Minimize Reconstruction Error: Ensures the decoder’s output closely matches the input data by minimizing reconstruction error. This process leverages the K-Means clustering algorithm to group data points into clusters based on their proximity to centroids.
‍

2) Enforce Latent Diversity and Prevent Overfitting: Regularizes the latent space by ensuring sampled latent vectors follow a normal distribution. This is achieved by minimizing the divergence loss (Kullback-Leibler Divergence), which quantifies the difference between two probability distributions.
‍

Result

Reconstruct Input Images

The images below show the decoder's output after 100 epochs of training (learning rate = 0.0007). The results demonstrate the VAE's ability to reconstruct street maps, capturing the general layout of the source images.

However, due to the limited dataset (592 images), the resolution is not optimal, and finer details of the source images are not fully revealed.

New Data Generation: Sampling from learned distribution in latent space 

We project both source and target images into the latent space through the trained encoder.

By calculating the mean vector of both latent vectors, we can interpolate between them and generate new images through the trained decoder.

The results show that the newly generated images inherit characteristics from both the source and target images. However, due to the small dataset, the images lack sufficient resolution to clearly display these inherited traits.

Source

Generated

Target

New Data Generation: Medieval Street Style

Interpolating the Latent Space within Medieval Styles

New Data Generation: Radial Street Style

Interpolating the Latent Space within Grid Styles

New Data Generation: Grid Street Style

Interploating the Latent Space within Grid Styles

New Data Generation: After-WW2 Street Style

Interploating the Latent Space within After-WW2 Styles

Contribution & Reflection

Model Performance v.s. Data Limitation

Our early exploration of generative urban street maps with variational autoencoders shows moderate success. While the model effectively reconstructed street maps, the small dataset size impacted both reconstruction quality and the generation of new images, resulting in less optimal outcomes.

Deep Learning in Architecture

The results underscore both the potential and challenges of applying deep learning in architecture. While deep learning offers creative possibilities beyond traditional tools, its black-box nature makes results unpredictable and difficult to control. This uncertainty highlights the gap between AI's artistic applications and its practical use in solving real-world design problems.

‍