TLDR: The paper introduces a novel method for automated facility enumeration in building compliance checking. It combines door detection with Large Language Models (LLMs) and a Chain-of-Thought (CoT) reasoning pipeline to accurately identify and count various facility types (like toilets, kitchens, exits, parking) from CAD floor plans. This approach significantly outperforms traditional LLM baselines, offering a robust and generalizable solution to accelerate design validation and reduce manual effort in meeting regulatory standards.
Ensuring that buildings meet regulatory standards, a process known as Building Compliance Checking (BCC), is crucial for safety, functionality, and sustainability. A key part of this process involves accurately counting and verifying the spatial distribution of various facilities within a building, such as sanitary facilities, kitchens, laundries, exits, and parking spaces. Traditionally, this task has been performed manually, making it time-consuming and labor-intensive.
Recent advancements in artificial intelligence, particularly Large Language Models (LLMs), are opening new doors for automation in this field. A new research paper titled Automated Facility Enumeration for Building Compliance Checking using Door Detection and Large Language Models by Licheng Zhang, Bach Le, Naveed Akhtar, and Tuan Ngo introduces a novel approach to automate facility enumeration, a task that has largely been overlooked in existing literature.
The Challenge of Manual Compliance
Manual building compliance checking is prone to errors and inefficiencies. Even a single miscount of a critical facility like an emergency exit or fire extinguisher can lead to serious safety and legal risks, invalidating a design and delaying project approval. The complexity arises from varying architectural standards, diverse design styles, and the need for contextual reasoning to ensure accurate compliance with regulations like Australia’s National Construction Code (NCC).
A Novel AI-Powered Solution
The researchers propose a method that integrates door detection with LLM-based reasoning, enhanced by a Chain-of-Thought (CoT) pipeline. This approach is designed to validate the quantity of each facility type against statutory requirements directly from CAD floor plan images. Unlike traditional vision-only methods, LLMs can combine visual understanding with textual interpretation, allowing them to generalize across different regulatory contexts and interpret complex building requirements.
How the System Works
The proposed pipeline involves several structured reasoning steps:
First, the system uses a state-of-the-art object detection model called Co-DETR to accurately identify all door instances in a floor plan image. Doors are chosen as anchor points because they have stable symbolic representations across different drawings and help the LLM focus on local regions.
Next, for each detected door, the LLM (specifically, OpenAI’s GPT-5) predicts the type of room it connects to. For example, it determines if a door leads to a toilet.
Following this, the LLM performs a “room consolidation” step. If multiple doors are predicted to lead to the same type of room, the LLM identifies which doors belong to the same room, eliminating duplicates to ensure each unique facility is counted only once.
Finally, an “omission correction” step is introduced. The LLM reviews the entire floor plan, with the already identified facilities highlighted, to detect any missing instances that might have been overlooked in the previous steps, including rooms without associated doors.
The final count for a specific facility type is then the sum of the consolidated doors and any newly identified missing instances.
Practical Benefits and Performance
This automated approach offers significant practical applications. It can automatically verify sanitary facilities, kitchens, laundries, exits, fire safety equipment, accessibility provisions, and parking with minimal human intervention. This accelerates design validation, reduces manual inspection efforts, and enhances the accuracy of facility enumeration across diverse building projects.
Experiments conducted on various real-world and synthetic floor plan datasets demonstrated the effectiveness and robustness of the method. The CoT pipeline consistently outperformed a baseline GPT-5 model across categories like toilets, kitchens, emergency exit doors, and car parking. Notably, encouraging the LLM to explicitly justify its predictions by appending “tell me the reason” to prompts consistently improved performance, suggesting that deliberate reasoning enhances accuracy.
Also Read:
- NUMINA: A New Benchmark Reveals AI’s Struggle with 3D Numerical Reasoning
- Mapping Indoor Spaces: A New Approach to Floorplan Reconstruction with AI
Conclusion
By introducing automated facility enumeration and leveraging the combined power of door detection and LLM-based Chain-of-Thought reasoning, this research provides a scalable and accurate solution for building compliance checking. This advancement has the potential to significantly transform compliance verification practices in the architecture, engineering, and construction (AEC) industry, making the process more efficient, reliable, and adaptable to diverse regulatory requirements.
