KOLCO CLUB · Spatial reasoning
300,000 floor plans.On real CAD layers.
Indian residential plans, studio to four bedroom, drawn as editable SVG with every room name, dimension, area figure and schedule row as live text — and every wall, door, column and fixture on its own standard architectural layer.
floor plans
median 1,123
every plan
and metadata
Not a picture of a plan. A plan.
Auto-traced floor plans give you one undifferentiated soup of paths. These are drawn on the AIA layer convention, so walls are separable from furniture, and annotation is separable from architecture, without segmenting anything first. All 50 plans in the public sample carry 20 of these layers; none carries fewer.
Room labels come as a regular quartet — name, metric dimensions, imperial dimensions, area — so room polygons, adjacency graphs and area regression targets are all recoverable from the drawing itself. Median 139 live text nodes and 275 path elements per plan, measured on the sample.
What 300,000 plans actually look like.
Stated up front, because you will compute it in your first evaluation hour. The distribution follows what actually gets built in India, which makes it realistic and unbalanced at the same time.
By unit type
By footprint shape
By carpet area
Three-bedroom plans and L-shaped footprints dominate, and areas cluster hard between 1,000 and 1,249 sq ft. If your model needs a flat distribution rather than a realistic one, a release can be sampled to a target balance, or the thin classes drawn to order.
Twenty-one fields per plan.
One CSV row per file, shipped with the pack. This is a real row, not a schema sketch.
- Programme Unit type, rooms, toilets, attached toilets, balconies
- Area Carpet area in square feet and square metres
- Geometry Footprint shape, circulation type
- Orientation Facing, vastu alignment
- Drawing Dimension system, palette, typeface, symbol style, detail level
- File Score and size in KB
What it is for.
Plan understanding
Room detection, area regression, unit-type classification and adjacency extraction, with labels that come from the drawing rather than from a human annotator’s guess.
Plan generation
Conditional layout generation — given a footprint, an area budget and a room programme, produce a plan. The metadata gives you the conditioning variables directly.
Raster to CAD
Every plan renders deterministically to an image, so each file is an aligned pair for training image-to-vector conversion on technical drawings rather than on logos.
Document and drawing VQA
Questions with checkable answers — how many bedrooms, which way does the master face, what is the balcony area — grounded in live text you can read without OCR.