Blueprint Segmentation with Mask R-CNN

The Challenge

Indoor mapping needed a way to automatically identify walls, doors, points of interest, and text directly from raw blueprint images.

What We Built

We annotated blueprint imagery in Roboflow and trained a Mask R-CNN model (via Detectron2) for instance segmentation, then converted blueprint coordinates into real-world coordinates using affine transformation.

• Custom-annotated blueprint dataset built in Roboflow for training and validation
• Mask R-CNN instance segmentation model trained via Detectron2 to detect walls, doors, and points of interest
• Text recognition to extract room labels and identifiers directly from blueprint images
• Affine transformation pipeline to convert pixel coordinates into real-world building coordinates
• Structured output feeding directly into the Mapify routing engine

Where It Stands

Shipped in one month as the computer-vision backbone powering the Mapify indoor navigation apps.