Annotating a Custom Dataset for Blueprint Segmentation in Roboflow

Good models start with good data, not better architecture

The Mapify indoor navigation apps depend on a model that can read a raw architectural blueprint and identify walls, doors, points of interest, and text automatically. Before any of that model architecture mattered, we needed a dataset that actually taught it what those things look like across different blueprint styles, and no public dataset for that existed.

Building the dataset by hand

We annotated a custom blueprint dataset in Roboflow, labeling walls, doors, and points of interest across a range of real architectural blueprints. This is unglamorous work compared to model training, but it is where most of the actual accuracy in a computer vision project gets decided. A model trained on inconsistent or thin annotations will find inconsistent, thin patterns, no matter how sophisticated the architecture is.

Training with Detectron2

With the annotated dataset in place, we trained a Mask R-CNN model via Detectron2 for instance segmentation, teaching it to distinguish individual walls, doors, and points of interest as separate objects rather than one undifferentiated shape. Text recognition ran alongside this to extract room labels directly from the blueprint images.

The output of this pipeline is structured data, not just a segmented image, which is what let it feed directly into the affine transformation step that converts blueprint coordinates into real-world building coordinates for the Mapify routing engine.

What this looked like as a standalone engagement

This was delivered in one month as the computer-vision backbone powering the Mapify apps. If you are scoping a similar computer vision project, budget real time for dataset annotation specifically. It is usually the least predictable part of the timeline and the most important one to get right.