House-GAN: Relational Generative Adversarial Networks for Graph-constrained House Layout Generation (2003.06988v1)

Abstract: This paper proposes a novel graph-constrained generative adversarial network, whose generator and discriminator are built upon relational architecture. The main idea is to encode the constraint into the graph structure of its relational networks. We have demonstrated the proposed architecture for a new house layout generation problem, whose task is to take an architectural constraint as a graph (i.e., the number and types of rooms with their spatial adjacency) and produce a set of axis-aligned bounding boxes of rooms. We measure the quality of generated house layouts with the three metrics: the realism, the diversity, and the compatibility with the input graph constraint. Our qualitative and quantitative evaluations over 117,000 real floorplan images demonstrate that the proposed approach outperforms existing methods and baselines. We will publicly share all our code and data.

• The paper introduces House-GAN, which integrates relational mechanisms into GAN architecture to generate house layouts that adhere to graph-based architectural constraints.
• It employs convolutional message passing networks (Conv-MPN) for effective spatial feature propagation, outperforming traditional methods in realism and diversity.
• Experimental results on 117,000 floorplan images validate House-GAN's superior compatibility and performance using metrics like FID and graph edit distance.

Relational Generative Adversarial Networks for Graph-Constrained House Layout Generation

The paper presents House-GAN, a novel approach in the domain of generative adversarial networks (GANs), specifically devised to create house layouts that adhere to predefined architectural constraints conveyed through a graph, referred to as a "bubble diagram." The proposed network architecture incorporates relational mechanisms into both the generator and discriminator, focusing on retaining the structural integrity and spatial relationships dictated by the input constraint graphs. The task at hand involves the generation of axis-aligned room bounding boxes based on the given architectural constraints.

Methodology Overview

House-GAN distinguishes itself by encoding architectural constraints directly into the graph structure of the neural networks used in the GAN's architecture. It utilizes a convolutional message passing network (Conv-MPN) to propagate information across graph-structured data, efficiently reasoning about the spatial and relational attributes necessary for generating viable house layouts. This mechanism stands in contrast to traditional graph convolutional networks (GCNs), as it operates in feature volumes rather than latent spaces, allowing for more effective feature transformations specific to design spaces.

The framework describes layout generation in three phases:

1. Input Graph Formation: This phase constructs a relational graph from the bubble diagram, initializing each node with room-specific features.
2. Conv-MPN/Upsampling: It employs convolutional message passing to update node features, followed by upsampling operations to maintain resolution.
3. Layout Output: The final feature volumes are decoded into binary room masks, which are thresholded to derive room bounding boxes.

In contrast, the discriminator processes input room segmentation masks to differentiate real layouts from synthetic ones. The relational architecture of the discriminator also involves Conv-MPN operations which leverage graph-based adjacency constraints.

Experimental Results

The authors evaluated the proposed model against several baselines and competing models utilizing a substantial dataset of 117,000 real floorplan images to benchmark realism, diversity, and compatibility of generated layouts:

1. Realism was assessed via a user paper, where professional architects and graduate students compared generated layouts with actual data, showing House-GAN to have superior user scores.
2. Diversity was measured through Fréchet Inception Distance (FID) scores, where House-GAN demonstrated the ability to generate variations while maintaining high-quality outputs, outperforming other methods in most categories.
3. Compatibility was gauged using graph edit distance metrics, attesting to House-GAN's proficiency in producing layouts that adhere closely to input bubble diagrams without sacrificing diversity.

Implications and Future Outlook

House-GAN's approach illustrates an advancement in utilizing relational networks for structured data tasks, notably in architectural design, highlighting its potential to automate the initial phases of house design under specific constraints. It succeeds in generating layouts that fulfill spatial adjacency requirements while maintaining room arrangement flexibility, crucial for patient-specific architectural uses.

Future explorations may involve refining the model to accommodate complex room shapes, incorporating door or window placements, and integrating functionalities for designing multi-level structures. These advancements could significantly enhance automated architectural design processes, making them more adaptable to real-world scenarios and complex architectural demands.