Floor-plan datasets for architecture AI

What exists

The established floor-plan corpora are research datasets. RPLAN and LIFULL HOME'S are large collections of Japanese apartment layouts. CubiCasa5K is a few thousand annotated Finnish plans. ResPlan is more recent and smaller. All of them have advanced the field, and all of them share the same three limits for a commercial team.

The three limits

• Licensing. Research use is granted; commercial model training generally is not. That is a hard stop for a product team, and it cannot be resolved after training.
• Geography. The layouts encode the residential conventions of the country they came from. A model trained on Japanese apartment plans has not learned Indian residential practice, or American, or European.
• Representation. Several are raster images with annotation layers rather than vector documents, so wall geometry, room polygons and dimension text have to be recovered by a detector instead of read.

What a vector plan gives you instead

Walls as geometry rather than dark pixels. Rooms as closed polygons with an area you can compute rather than estimate. Dimension strings as live text. Doors, windows and fixtures on their own layers, so a model can be asked about one class without segmenting the drawing first.

It also makes evaluation tractable: intersection-over-union over room polygons is a real metric when the polygons exist in the file.

Questions to ask a vendor

Which residential practice the plans follow, whether any real property is depicted, whether the geometry ships alongside the drawing, and how duplicates were measured. A plan set with no duplicate analysis is usually a plan set with duplicates.