Reading Architectural Blueprints with Computer Vision: The Mapify Pipeline

Why indoor navigation cannot rely on GPS

GPS does not work indoors, which leaves large buildings, campuses, and facilities without a reliable way to give visitors turn-by-turn directions the way outdoor maps take for granted. The manual alternative, hand-building a digital map for every building, does not scale past a handful of locations.

Mapify's answer was to let organizations upload the blueprint they already have and let a computer vision pipeline do the mapping work automatically.

Turning a static image into structured data

A blueprint is just an image until something can identify what is actually in it: walls, doors, points of interest, and room labels. We annotated a custom blueprint dataset in Roboflow and trained a Mask R-CNN model, run through Detectron2, to perform instance segmentation on architectural blueprints, picking out these elements individually rather than treating the image as one flat picture.

Text recognition runs alongside the segmentation model to extract room labels and identifiers directly from the blueprint image, so a room does not just get detected as a shape, it gets identified by name.

From pixels to real-world coordinates

Detecting a wall in an image is only useful if the system knows where that wall actually sits in physical space. We built an affine transformation pipeline that converts blueprint pixel coordinates into real-world building coordinates, which is what makes accurate turn-by-turn routing possible instead of a map that merely looks right.

That structured output feeds directly into the Mapify routing engine, which calculates the shortest path between any two points of interest inside a building in real time.

Two apps built on one pipeline

The computer vision pipeline is the backbone, not the whole product. One pipeline supports onboarding multiple organizations and buildings from a single manager account, which is what let this scale past a one-off project.

• Manager app: registering organizations and uploading building blueprints
• End-user app: searching destinations and getting turn-by-turn indoor directions