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

Area of Science:

• Computational Architecture
• Artificial Intelligence in Design
• Human-Computer Interaction

Background:

• Automated floor plan generation is crucial for efficient architectural design.
• Existing methods often lack human-centric considerations, leading to suboptimal spatial arrangements.
• Integrating human activity and preferences into design processes remains a challenge.

Purpose of the Study:

• To introduce a novel two-stage, human-centric approach for automated residential floor plan design.
• To enhance the geometric plausibility and topological reasonableness of generated floor plans.
• To improve resident interaction with the designed architectural environment.

Main Methods:

• Synthesizing a human-activity map from an exterior wall boundary using generative adversarial networks (GANs) or semi-automated user input.
• Employing a deep framework, ActFloor-GAN, with a cycle-consistency constraint for pixel-wise room type prediction.
• Converting predicted room layouts into vectorized floor plans.

Main Results:

• Quantitative comparison demonstrates superior performance in predicting room types by leveraging the human-activity map.
• Subjective evaluation by architects indicates compelling quality comparable to professional designs.
• Generated floor plans exhibit significantly improved room layout topology over existing methods.

Conclusions:

• The proposed human-centric approach significantly enhances automated floor plan generation.
• Incorporating human activity maps and user preferences leads to superior, more interactive architectural designs.
• The method offers a powerful tool for creating high-quality, user-centered residential floor plans.