Case study / Mar.2025 / Dev & AI / Client: BTP Consultants
Floorplan Computer Vision
Floorplan Computer Vision was an R&D project to explore how to detect and extract polygons from floorplans with enough quality for downstream compliance workflows. The key goal was to reduce dependency on Kreo and prepare support for new plan types (electrical, facade, plumbing). I delivered the project solo: internal-plan ingestion pipeline, outsourced annotation workflow through Roboflow, dataset building, model selection, training file generation, SageMaker training runs, and inference pipeline deployment. The inference architecture used ALB-based scaling to control costs and reduce cold-start impact. The project established a reusable method for dataset creation, training, and inference that is now reused for new plan domains, while room-mask quality constraints were handled through a more precise and cost-efficient room extraction technique based on walls (now used in AnalyzTech).
01 / Brief
Solution
Floorplan Computer Vision was an R&D project to explore how to detect and extract polygons from floorplans with enough quality for downstream compliance workflows. The key goal was to reduce dependency on Kreo and prepare support for new plan types (electrical, facade, plumbing). I delivered the project solo: internal-plan ingestion pipeline, outsourced annotation workflow through Roboflow, dataset building, model selection, training file generation, SageMaker training runs, and inference pipeline deployment. The inference architecture used ALB-based scaling to control costs and reduce cold-start impact. The project established a reusable method for dataset creation, training, and inference that is now reused for new plan domains, while room-mask quality constraints were handled through a more precise and cost-efficient room extraction technique based on walls (now used in AnalyzTech).
02 / Constraints
Problem
- Handling very high-resolution plans (~6000x6000) while preserving segmentation quality
- Achieving surgical mask precision for room boundaries, not only good classification
- Balancing annotation effort, infrastructure cost, and model performance
- Building robust training and inference workflows reusable across future plan domains
03 / System
Architecture / workflow
- Solo end-to-end R&D delivery: dataset pipeline, training, inference, and deployment
- Reduced dependency on Kreo and prepared extension to new plan types
- Reusable method for CV dataset creation, training, and inference for floorplan domains
- ALB-based inference scaling to optimize cost while limiting cold-start impact
- Transfer of learning into AnalyzTech through a more precise room extraction approach
04 / Evidence
Results / impact
- Furniture detection reached 88 mAP@50
- Room detection reached 60 mAP@75
- Room masks were not accurate enough for production-grade usage
- Established a reusable method for dataset creation, training, and inference
- Enabled follow-up datasets for new plan types (electrical, plumbing)