Loading
Research · Thread 10
Floor-plan extraction
Six pipeline generations, scored on the same eighteen scans, and the honest result: the simplest one won.
A scan already knows the shape of a building, but a buyer wants a floor plan: walls, rooms and doors, drawn to scale on one page. The hard part is deciding what counts as a wall, when wardrobes, curtains, kitchen islands and half-height partitions all read as vertical surfaces to a point cloud. Six pipeline generations were built, from fitting geometry directly to the point cloud, through grid and footprint methods, to a learned model built on published architectural-reconstruction research, and every one was scored against the same eighteen scans rather than demoed on the ones it handled well.
Finding
The simplest method won
Six pipeline generations, geometry-fitting through to a learned model, scored on the same eighteen scans. The learned approach is the most sophisticated and won only two of eighteen, three as an ensemble: it learns from synthetic buildings and comes out 30 to 80 percent smaller on real scanned ones.
A gap between training data and deployment, not a tuning problem, and it did not close with more tuning.
Paper to shipped
- Six generations, kept side by side
Each new generation stayed runnable rather than replacing the last, geometry-fitting through grid and footprint methods to a learned model, which is what made it possible to find out later that a newer one was worse.
- The learned model lost, and kept losing
The machine-learning pipeline is the most sophisticated approach here and wins on two scans out of eighteen, three once it is paired with a plain method's building outline. It was not dismissed on a first result: the training scans were auto-labelled, it went through two rounds of fine-tuning, the scoring was overhauled, and the output polygons were expanded. Its scores improved substantially; the win rate did not move until it was run as an ensemble.
- Why it loses, and what ships instead
It learns from synthetic buildings and is asked to draw real scanned ones, coming out 30 to 80 percent smaller than the true area, a gap between training data and deployment rather than a tuning problem. The apartment plan published on this site today is drawn by the ninth version, a plain grid method, labelled grid_v9 in its own title block, and the villa beside it by v11 footprint. The scoring function is its own bottleneck too: it rewards rectangularity and low room count, which is what the grid methods already produce. Next: pairing the learned model with something that gets the outer shape right, and fixing the scoring.
I'm Pavlo Puzikov, a marketer working in immersive 3D, applied AI, and generative content.