Scan to Floor Plan: how it came together.
How it came together.
- 01
Six generations, kept side by side
The pipeline went through six approaches, from fitting geometry directly to the point cloud, through grid and building-footprint methods, to a learned model built on published architectural-reconstruction research. Six is the number of generations in this comparison, not a version number: the source repo labels its pipelines v7 through v15, so the third of the six is the one whose files and drawings read grid_v9. Each new generation stayed runnable rather than replacing the last, which is the decision the rest of this depends on: keeping the old ones alive is what made it possible to find out that a newer one was worse.
Data Pipelines - 02
Score them against each other, on the same scans
Every generation was scored on the same eighteen scans rather than demoed on the ones it handled well. That sounds obvious and is the step most often skipped, because the newest approach usually arrives with a story about why it is better and a couple of examples that support it. A shared scoreboard is what turns that story into a claim that can be wrong.
Research - 03
The learned model lost, and kept losing
The machine-learning pipeline is the most sophisticated thing here and it wins on two scans out of eighteen. 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 and the win rate did not move until it was run as an ensemble, paired with a plain method's building outline, which took it to three. Being beaten after a fair fight is a much more useful result than being beaten because nobody tried.
AI & Machine Learning - 04
Why it loses, which is the part worth keeping
The model learns from synthetic buildings and is asked to draw real scanned ones, and on real scans its rooms come out consistently smaller than the rooms actually are, by a margin large enough to lose on total area even where it identifies the layout correctly. That is a gap between where the training data comes from and where the work happens, not a tuning problem, which is why more fine-tuning kept not fixing it and why the next thing to try is combining it with a method that gets the outer shape right.
3D & Spatial - 05
The scoreboard had a bias of its own
The scoring function that produced every number above rewards rectangularity and a low room count, which is a description of what the grid methods already make. It pays a plain four-vertex rectangle more than an architecturally correct polygon, so the learned model was marked down for drawing shapes closer to the real building. My own notes name the scoring function as the real bottleneck. This does not overturn the result, because the plain methods also produce plans a client can read, but the race was partly refereed in the winner's favour, and the next honest step is to fix the referee rather than tune the runners again.
Research