The Problem: Floor Plans Are Locked in PDFs
Real estate agencies, architecture firms, and property managers deal with floor plans every day. But the data inside those floor plans — room dimensions, square footage, layout relationships — is trapped in images and PDFs.
To get structured data, someone has to:
- Manually measure each room from the drawing
- Calculate square footage room by room
- Type it all into a spreadsheet or listing platform
- Double-check because one wrong measurement means wrong pricing
For a real estate agency listing 50 properties per month, that's hours of tedious work per listing. For an architecture firm reviewing plans, it's even worse.
What if AI could look at a floor plan and instantly tell you every room, its dimensions, and total area?
What We Built
We created an AI system that takes any floor plan image — scanned, photographed, or exported from CAD — and extracts structured spatial data.
How It Works
1. Image Analysis The floor plan is fed to our computer vision pipeline. We use a combination of line detection (for walls) and contour extraction (for room boundaries) to understand the spatial layout.
2. Room Detection The AI identifies individual rooms by detecting enclosed spaces formed by walls. It reads room labels when present, and infers room types (kitchen, bathroom, bedroom) from fixtures and layout cues when labels are missing.
3. Dimension Extraction Dimension annotations on the plan are read and matched to the corresponding walls. The system understands notation like "3.50m", "11'-6"", and metric/imperial mixed formats.
4. Structured Output Every room becomes a data object:
| Field | Example |
|---|---|
| Room Name | Living Room |
| Width | 5.2m |
| Length | 4.8m |
| Area | 24.96 m² |
| Doors | 2 (to Hallway, to Kitchen) |
| Windows | 3 |
| Floor | Ground |
Plus total property area, room count, and layout adjacency graph.
What Made This Challenging
Inconsistent Drawing Styles
Every architect draws differently. Some plans are minimal line drawings, others are richly detailed with furniture and textures. We needed a model that handles both extremes.
Scale and Units
Floor plans come in metric, imperial, and sometimes with no scale at all. We built a calibration step that detects the scale bar or known reference dimensions to normalize all measurements.
Multi-Floor Plans
Some documents contain multiple floors on a single page. The system detects floor boundaries and processes each independently.
Results
| Metric | Manual Process | Our System |
|---|---|---|
| Time per floor plan | 15-30 minutes | 10 seconds |
| Room detection accuracy | — | 95% |
| Dimension accuracy | Human error ~3% | ±2% variance |
| Output format | Spreadsheet | JSON + visual overlay |
For a real estate platform processing 200 listings/month:
- Before: 2 staff members spending 3 hours/day on floor plan data
- After: Automated processing with 10 minutes of daily review
- Result: Listings go live 2 days faster, with more accurate square footage data
Who This Is For
- Estate agents who need room sizes for listings without measuring
- Property managers turning a shelf of old plans into a searchable list
- Architects pulling room schedules out of older drawings
- Interior design tools that need room shapes to plan a space
- Builders turning as-built drawings into data
