Generating 2D Floor Plan Using Deep Learning

Posted by Eric on October 29, 2024

This is a project to explore how to generate 2D floor plan using different methods.

Archi-GAN

The 1st method I tried is called Archi-GAN, an method posted in 2019. Demo

Dataset

TELL2DESIGN–This is from a research <Tell2Design: A Dataset for Language-Guided Floor Plan Generation> in 2023. Demo Demo

Process

Load Data –> Resize & Pre-process –> Split to train & test sets –> Train Model (/w 200 epoches & 500 epoches) –> Predict

Test Record

Demo Demo With 200 epoches, it cost 250 minutes. With 500 epoches, it cost 700 minutes. (Maybe I need a faster computer)

Generated Floor Plan

Demo Demo Demo

As shown in the outputs, the results are really poor :( I’ll try some other methods later.

A Big Step Forward! ;)

GAT-NET

I tried the project codes of GAT-NET, and successfully generated floorplan.

Architecture

Demo

It looks a little bit complicated. But the idea is very clear. Firstly, the centroids are estimated and then the boundaries of the spaces are generated.

Test Case

Demo

This is the original layout.

Demo

Here is the boundary.

Demo

Here is the centroids which are generated by CNN. I will find out how does this part work:)

Demo

Done! The generated floorplan is shown. It looks not bad, at least reasonable from my perspective of view.

Lots of question marks are in my mind. But, it is a good start!