SUCCESS STORY
Case Study: Automated Data Extraction of Set Products from Floor Plans

Automated Data Extraction of Set Products from Floor Plans

ThirdEye Data successfully delivered a Proof of Concept (PoC) that used state-of-the-art object detection models to identify and structure predefined architectural objects from image-based floor plans. The MVP phase is currently under discussion following the successful completion of the PoC.

THE CUSTOMER

The customer is a large scale enterprise solutions provider, specializing in construction-related automation, supply chain workflows, and CRM integrations. Their primary client is a top-tier architectural and structural metal manufacturer. The customer has a big in-house design team to create and modify detailed building layouts.

BUSINESS GOALS OR CHALLENGES

Business Goals

  • Automate the extraction of predefined construction set products from unstructured floor plan images.
  • Convert visual design data into structured digital output compatible with tools like ArchiCAD.
  • Reduce reliance on manual annotations and accelerate the drafting process.
  • Enable scalable, intelligent design review and object detection workflows.

Understanding the Challenges:

  • Floor plans existed only as static images, without vector or layer data.
  • Manual detection of embedded objects was labor-intensive and error-prone.
  • Wide variation in complexity and object positioning across 4,000+ plans.
  • Required high accuracy in detecting multiple object types and aligning with wall layouts.
  • Needed integration with downstream tools used by internal design teams.

Prerequisites and Preconditions:

  • Access to a large dataset of real-world floor plan images.
  • Manual annotations using tools like Roboflow for training data preparation.
  • Pre-agreed success benchmarks for object detection across simple to complex layouts.
  • Approval of bounding box visualizations and JSON output structure.
  • Readiness to iteratively improve detection accuracy based on real-world feedback.

THE SOLUTION

ThirdEye Data delivered a PoC project, focusing on building and optimizing a deep learning pipeline using YOLOv8 for object detection and OpenCV for image post-processing.

Solution Highlights

  • Developed a deep learning-based multi-object detection system using YOLOv8 to identify and quantify key interior construction products in floor plan images.

  • Integrated OCR to extract specification codes related to doors, windows, and fixtures.

  • Applied image preprocessing and post-processing logic using OpenCV and Pillow for enhanced boundary detection and noise reduction.

  • Generated structured output in JSON format, capturing object class, coordinates, and count for each detected object.

  • Enabled UI-based image uploads and red-box visual feedback for easy validation by design teams.

  • Built a modular and scalable backend inference engine using FastAPI, containerized for production deployment.

  • Prepared datasets using Roboflow for manual annotation, version control, and augmentation to improve model generalization.

Technologies Used

  • Object Detection & Training

    • YOLOv8 for real-time multi-object detection

    • Detectron2 as an alternative benchmarking model

    • PyTorch for deep learning pipeline and model training

  • Image Preprocessing & Alignment

    • OpenCV for edge detection, noise reduction, and alignment

    • Pillow for image manipulation and post-processing

  • Text & Code Extraction

    • EasyOCR and Tesseract for extracting window and door codes

    • spaCy for parsing text annotations and specification tags

  • Data Annotation & Management

    • Roboflow for image annotation, versioning, and augmentation

    • Label Studio for team-based labeling workflows

  • Backend & Integration

    • FastAPI for serving inference results via REST API

    • Docker for containerized deployment

    • ezdxf and AutoCAD .NET API for DXF file creation and CAD integration

  • UI & Visualization

    • Streamlit and Dash for visualizing floor plan inputs and red-box outputs

    • Matplotlib for bounding box overlays and object summaries

  • Data Output Format

    • JSON for structured results (object class, coordinates, count)

    • DXF for CAD-compatible drawing exports

VALUE CREATED

  • 80–90% detection accuracy achieved for simple and mid-complex floor plans.
  • Over 4,000 plans processed, drastically reducing manual annotation workload.
  • 65% time savings in identifying and placing set products within floor plans.
  • Improved detection consistency across over 5 distinct object categories.
  • Delivered structured JSON and DXF outputs, enabling smooth downstream integration with ArchiCAD and AutoCAD tools.
  • Enabled real-time validation of detected results using visual red-box overlays, improving design team productivity.
  • Reduced average design review cycle from 2–3 days to a few hours for batch uploads.
  • Positioned the customer to scale the system across all incoming architectural plans with minimal human intervention.
ThirdEye Data

Transforming Enterprises with
Data & AI Services & Solutions.

ThirdEye delivers Data and AI services & solutions for enterprises worldwide by
leveraging state-of-the-art Data & AI technologies.

Talk to ThirdEye