ActFloor-GAN: Activity-Guided Adversarial Networks for Human-Centric Floorplan Design

Area of Science:

• Computer Science
• Artificial Intelligence
• Architecture

Background:

• Automated floorplan generation is crucial for efficient residential building design.
• Existing methods often lack human-centric considerations, leading to suboptimal spatial layouts.
• Integrating human activity and environmental interaction is key to improving architectural space usability.

Purpose of the Study:

• To develop a novel two-stage, human-centric approach for automated floorplan design.
• To generate geometrically plausible and topologically reasonable residential floorplans.
• To enhance resident interaction with the architectural environment through intelligent design.

Main Methods:

• Synthesizing a human-activity map from an input boundary using generative adversarial networks (GANs) or semi-automatic user manipulation.
• Employing a deep framework, ActFloor-GAN, to predict room types based on the human-activity map.
• Utilizing a re-formulated cycle-consistency constraint within ActFloor-GAN to optimize prediction performance.

Main Results:

• The human-activity map significantly improves piecewise room type prediction compared to prior methods.
• Generated floorplans demonstrate compelling quality, comparable to professional designs, with superior room layout topology.
• The approach successfully incorporates user-design preferences by considering furniture placement and human activities.

Conclusions:

• The proposed two-stage, human-centric approach offers a significant advancement in automated floorplan design.
• ActFloor-GAN effectively generates high-quality, convertible vectorized floorplans that are both functional and aesthetically pleasing.
• This method enhances architectural design by prioritizing human interaction and user-specific preferences.