Floor model it33b (closed weights)
it33b turns a vector construction floor plan (PDF) into room floor polygons for flooring takeoff. It reads the drawing's own vector lines, decides which ones are walls, seals door openings, and builds rooms wall-face to wall-face. It is the current frozen "keep" version (frozen 2026-10-15) of Kentucky AI's September floor-model line.
The weights and code are closed. This repository is the model card only. Weights, code, and checksums are kept in a private repository and in our experiment tracker; access is by arrangement with Kentucky AI.
How it works
| Stage | Component | What it does |
|---|---|---|
| Picture | DeepLabV3-ResNet50 (3 classes), tiled | floor / wall / background probability on the rendered sheet |
| Wall lines | gradient-boosted classifier (scikit-learn) | scores every PDF vector segment as wall face or not |
| Doors | RF-DETR detectors (doors, door marks) | door leaves and swings -> hinge / strike points |
| Doorways | gradient-boosted classifier + size rules | picks jamb-to-jamb closures that seal each opening |
| Assembly | vector arrangement + rules | faces between wall lines and seals -> rooms; fixture holes filled; out-of-scope views removed |
What changed from the previous version (it32): one rule that deleted large open floors (gyms, open offices) was removed, and the view-title reader now recognises titles without the word "PLAN", so a plan is not mistaken for an enlarged detail or a reflected ceiling plan. Reflected-ceiling, enlarged, elevation, and roof views get no floor by design.
Evaluation (measured, 2026-10-15)
The 51 reviewed sheets are scored with wall-line and doorway classifiers refit without each sheet's job (job hold-out); the 33 and 125 sheets come from jobs that were not used to train the line or doorway models. "Rooms at IoU 0.90" counts reference rooms matched by one predicted room overlapping it at least 90 %; it is a strict, per-room measure.
| Set | Result |
|---|---|
| 51 reviewed sheets (expert reference outlines) | 91.7 % of reference floor area found, 6.1 % extra; 71 % of rooms at IoU 0.80, 57 % at IoU 0.90; 54 % of rooms within 5 % of reference area |
| 33 unseen sheets from new jobs | 664 / 1,404 rooms at IoU 0.90 (identical to it32 apart from one sheet) |
| 125 held-out sheets from new jobs | 1,079 / 2,728 rooms at IoU 0.90 (it32: 1,070) |
| Control sheets | no regression vs it32; one gym control recovers 21,000+ SF of previously missed floor |
Known limitations
- Room edges can follow door-swing arcs, floor-finish hatch strokes, or dimension lines where these are scored as walls.
- Whole rooms are sometimes missed (left as no floor) on large multi-unit sheets.
- Merged rooms at doorways that were not sealed.
- Raster or scanned drawings are out of scope; the model needs vector PDFs with a known scale.
- It is a takeoff assistant, not a certified measurement system; outputs are meant to be reviewed.
Data
Trained and evaluated on private commercial and residential construction drawings with expert-reviewed floor outlines. No drawings, labels, or project data are distributed.
Licenses of components
DeepLabV3 / ResNet-50 architecture from torchvision (BSD-3-Clause); RF-DETR (Apache-2.0); scikit-learn (BSD-3-Clause). The trained weights are proprietary to Kentucky AI and are not licensed for redistribution.