RAQEEB-CCTV v1.0.0

Floor-plan element segmentation for CCTV planning. One YOLO11s-seg model that finds and outlines eleven building elements on architectural floor plans: walls, doors, windows, open passages, columns, stairs, lifts, escalators, exit markers, gates and rooms/spaces. Each detection is a polygon mask with a bounding box, a class name and a confidence. The model works on raster plans (PNG, JPEG, scanned or exported PDF), and a companion tool runs it on AutoCAD drawings (DWG/DXF) and writes the labelled boxes back into the drawing.

It was trained on 62,205 labelled plan images merged from 16 public sources, with a partial-label training scheme so that datasets which only annotate some classes do not teach the model that the other classes are background.

RAQEEB-CCTV on a held-out architectural plan

Version 1.0.0 (2026-11-14)
Architecture YOLO11s-seg, 10.1 M parameters, 1024 px input
Weights raqeeb-cctv-v1.0.0.pt (20.6 MB, fp16) + thresholds.json
Held-out test mask mAP50 0.691, mask mAP50-95 0.427, box mAP50-95 0.567 (1,682 images)
AutoCAD check real DWG, held-out floors: doors 96-99% found; stairs/rooms partly; walls, windows, columns not reliable (details)
Licence CC BY-NC 4.0 (non-commercial; inherited from CubiCasa5K and FloorPlanCAD)

Classes

id class what it marks test mask mAP50 status
0 wall wall thickness, cut at openings 0.530 usable
1 door doorway / portal footprint 0.820 strong
2 window glazing / opening along a wall 0.636 usable
3 open_passage opening without a door 0.434 weak (few examples)
4 column structural column / pillar 0.300 weak
5 stair stair flights and landings 0.761 strong
6 elevator lift car / shaft 0.974 strong
7 escalator escalator footprint 0.606 (4 instances) not validated
8 exit_marker printed exit / egress sign 0.952 strong
9 gate pedestrian / vehicle gate — not supported (no training data)
10 space_region room, corridor or defined zone 0.898 strong

Quick start

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download("swedishtech/RAQEEB-CCTV", "raqeeb-cctv-v1.0.0.pt", revision="v1.0.0")
model = YOLO(weights)
result = model.predict("plan.png", imgsz=1024, conf=0.25)[0]   # whole plan, no tiling
for box, cls, conf in zip(result.boxes.xyxy, result.boxes.cls, result.boxes.conf):
    print(model.names[int(cls)], round(float(conf), 2), box.tolist())
result.save("plan_detections.jpg")

For best results use the per-class operating thresholds in thresholds.json (chosen on the validation split) instead of one global conf. The scripts in code/ do this for you (pip install -r code/requirements.txt):

# labelled boxes + before/after image for PNG / JPEG / PDF pages
python code/annotate_raqeeb.py plan.pdf --weights raqeeb-cctv-v1.0.0.pt --thresholds thresholds.json --out out/
# polygons in page coordinates + review queue (JSON) and overlays
python code/infer_raqeeb.py plan.pdf --weights raqeeb-cctv-v1.0.0.pt --thresholds thresholds.json --output out/
# AutoCAD drawing: labelled boxes on the drawing and back inside a copy of the DWG (one layer per class)
python code/cad_raqeeb.py drawing.dwg --weights raqeeb-cctv-v1.0.0.pt --thresholds thresholds.json --out out/ \
    --region floor1=x0,y0,x1,y1 --px-per-unit 0.1 --plot

Use whole pages for raster plans. The model was trained on whole plans resized to 1024 px. Cutting a 300 dpi page into 1024 px tiles makes rooms about three times larger than anything seen in training and produces false, tile-shaped room regions; infer_raqeeb.py and annotate_raqeeb.py therefore default to whole-page inference. CAD drawings are the exception: they are rendered at a fixed drawing scale and tiled (see below).

Results

Held-out test split

1,682 images whose drawings never appear in training (splits are grouped by near-duplicate drawing). Each image is scored only on the classes its source fully annotates (see Partial labels). conf 0.001, max_det 1500, 1024 px.

class instances images mask P mask R mask mAP50 mask mAP50-95 best F1
wall 30,469 1,061 0.620 0.513 0.530 0.230 0.564
door 7,579 893 0.843 0.737 0.820 0.488 0.789
window 7,326 652 0.703 0.664 0.636 0.257 0.685
open_passage 47 23 0.516 0.298 0.434 0.140 0.449
column 154 36 0.621 0.247 0.300 0.072 0.379
stair 426 225 0.737 0.736 0.761 0.511 0.751
elevator 142 75 0.898 0.965 0.974 0.833 0.939
escalator 4 3 0.436 0.970 0.606 0.507 0.615
exit_marker 76 24 0.804 0.974 0.952 0.526 0.943
gate 0 0 — — — — —
space_region 2,032 136 0.865 0.863 0.898 0.705 0.867
all 48,255 1,682 0.691 0.427

Walls and windows are long, thin shapes, so the strict mask mAP50-95 stays low even when their boxes are right (wall box mAP50-95 is 0.38 on validation against 0.21 for the mask).

Validation split (2,252 images): mask mAP50 0.616, mask mAP50-95 0.351, box mAP50-95 0.477 (metrics/val_METRICS.txt).

Against the project's previous model (same test images, same scoring; only 5 of the 1,682 images come from drawings that model saw in training):

model mask mAP50 mask mAP50-95 box mAP50-95
RAQEEB-CCTV v1.0.0 0.691 0.427 0.567
previous 11-class research model (334 training views) 0.108 0.055 0.059

Before and after: held-out test images

Every image below is from the test split; they were chosen by which classes they contain, not by how well the model did. "Correct" means a predicted box overlaps a ground-truth box of the same class with IoU >= 0.5 (one-to-one), counted only for the classes that source annotates (metrics/gallery_check.json).

Architectural plan (CubiCasa5K) — CubiCasa5K (high_quality_architectural/416), CC BY-NC 4.0. Correct / ground truth: wall 37/41 (42 predicted), door 9/10 (10 predicted), window 8/11 (9 predicted), open_passage 1/1, column 7/7 (10 predicted), stair 1/1, space_region 12/16 (14 predicted).

Architectural plan (CubiCasa5K)

CAD sheet (FloorPlanCAD) — FloorPlanCAD (CAD drawing crop), CC BY-NC 4.0. Correct / ground truth: door 18/19 (21 predicted), window 0/13, stair 1/1, elevator 0/1 (1 predicted).

CAD sheet (FloorPlanCAD)

CAD sheet with escalator (FloorPlanCAD) — FloorPlanCAD (CAD drawing crop), CC BY-NC 4.0. Correct / ground truth: door 5/5 (6 predicted), escalator 0/1.

CAD sheet with escalator (FloorPlanCAD)

Lifts, stairs and spaces — Roboflow Universe kt/floor-spac, CC BY 4.0. Correct / ground truth: stair 2/2, elevator 2/2, space_region 19/29.

Lifts, stairs and spaces

Evacuation plan with exit signs — Roboflow Universe tw/evacuation-plan, CC BY 4.0. Correct / ground truth: exit_marker 9/9.

Evacuation plan with exit signs

Working-drawing sheet — Roboflow Universe Doxle/floor-plan, CC BY 4.0. Correct / ground truth: wall 16/20 (18 predicted), door 7/9 (8 predicted), window 9/10 (11 predicted), stair 1/1.

Working-drawing sheet

Residential plan — Roboflow Universe Floorplanes/training-ai, CC BY 4.0. Correct / ground truth: wall 14/26 (19 predicted), door 2/9, window 5/5.

Residential plan

Simple floor plan — Roboflow Universe simple-floor-plan/floor-plan, CC BY 4.0. Correct / ground truth: door 6/17 (10 predicted), window 1/21 (6 predicted).

Simple floor plan

AutoCAD drawing check

Checked on a real AutoCAD drawing: the teaching-building construction drawing (DWG, drawn with TArch) from Men (2021), CC BY 4.0. code/cad_raqeeb.py opened the DWG with AutoCAD 2026 (headless accoreconsole), rendered each floor at 0.1 px per mm (100 px per metre, the scale of the CAD training images) in 1024 px tiles, ran the model with the operating thresholds, drew every detection as a box with its class name, and wrote the boxes back into the drawing: 597 detections became 1,194 new entities (a rectangle and a class confidence text each) on layers RAQEEB-DOOR, RAQEEB-WALL, ... AutoCAD then plotted the annotated drawing to PDF; the plot images below are AutoCAD's own output. The annotated drawing is in examples/autocad/ (DXF, opens in AutoCAD).

Because the drawing was made with TArch, its own objects are a ground truth: door, window and column blocks, the WALL and STAIR layers and the room-name texts. The render scale was chosen on the second floor (metrics/cad_verification.json); the ground and third floors below were not used for any choice.

class (truth = the drawing's own objects) ground floor third floor
door: found / in drawing 88 / 92 95 / 96
door: detections that are real doors 65% 77%
stair: found / stair groups 4 / 18 6 / 10
room-name texts inside a predicted space 21 / 62 15 / 66
window: found / in drawing 4 / 47 5 / 61
column: found / in drawing 0 / 119 0 / 106
WALL-layer lines covered by a wall mask 4% 3%

"Found" means the detection box covers at least half of the smaller of the two boxes (TArch door blocks include the swing arc, many training labels do not; with strict box IoU >= 0.5, 52 and 51 doors match).

Verdict for AutoCAD drawings (v1.0.0): the pipeline works end to end, and boxes and class names appear in AutoCAD on their own layers. On this drawing the model reliably marks doors, partly marks stairs and rooms, and does not reliably mark walls, windows or columns drawn in thin-line CAD style. Treat CAD output for those classes as unsupported in this version. (Two small re-rendered crops of this building, a stairwell and a freight lift, were in the training data; the evaluated floors were not.)

AutoCAD's plot of the annotated drawing (third floor, detail):

AutoCAD plot with RAQEEB boxes

Same area, rendered drawing before and after:

AutoCAD detail before and after

Whole ground floor (labels readable when zoomed in):

AutoCAD ground floor before and after

Intended use and limits

Intended use: research and non-commercial pre-processing of floor plans for CCTV / security planning: finding doors, stairs, lifts, exits and rooms so that a person can plan camera coverage faster. Every output should be reviewed by a person; the inference scripts put low-confidence and unvalidated detections in a review queue.

Not suitable for: commercial use (licence), safety-critical decisions such as certifying emergency exits, or any use without human review.

Known limits:

  • gate has no labelled training data anywhere in the inventory (4 instances); treat gate output as absent.
  • escalator has 360 training instances (all from FloorPlanCAD) and only 2 validation / 4 test instances; its threshold is unvalidated and every escalator detection goes to review.
  • column and open_passage are weak (few or inconsistent examples).
  • AutoCAD / CAD drawings: see the check above. Doors and stairs transfer well; CAD walls drawn as thin double lines, CAD windows and small solid columns are largely missed, because the CAD training source (FloorPlanCAD) was converted only for doors, windows, stairs, lifts and escalators, and the wall/column/room knowledge comes from raster plans.
  • Sources disagree on conventions (e.g. some Roboflow sets draw wall boxes straight across door openings); the model inherits some of that inconsistency.
  • The model detects symbols and footprints; it does not decide which doors are emergency exits, entrances or interior/exterior. Those need schedules, signage or a reviewer.

Training data

All sources listed in the project inventory that carry labels. Images were deduplicated and split by perceptual-hash drawing groups so near-duplicates of one drawing never straddle train and test.

source licence labelled classes used (fully annotated) train val test
CubiCasa5K CC BY-NC 4.0 wall, door, window, open_passage, column, stair, space_region 4,799 120 81
FloorPlanCAD CC BY-NC 4.0 door, window, stair, elevator, escalator 15,113 300 250
Floorplanes / training-ai CC BY 4.0 wall, door, window 11,252 300 250
Doxle / floor-plan CC BY 4.0 wall, door, window, stair 3,910 290 241
Floorplan-Segmentation CC BY 4.0 wall, door, window (+ rooms, partial) 5,306 48 29
kt / floor-plan CC BY 4.0 wall, door, window 1,213 181 67
Segmenting a Floor Plan CC BY 4.0 door, window (+ rooms, partial) 1,499 39 31
Floorplanes / walls Public Domain wall 1,704 256 209
Floorplanes / doors-detection2 CC BY 4.0 door 4,236 300 250
Floorplanes / doors-detection CC BY 4.0 window 3,997 2 1
kt / floor-spac CC BY 4.0 stair, elevator, space_region 963 96 43
ArchitecturalBlueprints CC BY 4.0 column (+ walls, partial) 99 10 4
tw / evacuation-plan CC BY 4.0 exit_marker (+ stairs, partial) 635 54 27
Testing / floor-plan-walls CC BY 4.0 wall 6,969 225 178
simple-floor-plan / floor-plan CC BY 4.0 door, window (+ rooms, partial) 198 21 9
earlier project review set (CubiCasa5K-derived plus a few CC BY sources) CC BY-NC 4.0 / CC BY 4.0 all 11 classes 312 10 12
total 62,205 2,252 1,682

Excluded: a 60-image "columns" set that turned out to be wine-rack photographs, and 2,402 images with no annotations at all. FloorPlanCAD drawings were rendered from the vector data at 1536 px and their symbol primitives grouped into instance polygons; CubiCasa5K SVG annotations were converted to polygons (walls cut at openings). Per-class training instances: wall 1,075,001 · door 279,157 · window 283,206 · space_region 99,286 · stair 13,479 · column 11,808 · elevator 4,360 · open_passage 2,330 · exit_marker 2,128 · escalator 360 · gate 4 (data/dataset_summary.json).

Training method

Merging 16 datasets that each label a different subset of the classes needs care: a doors-only dataset contains unlabelled walls, and naive training would teach the model that those walls are background.

  • Partial labels. data/label_sets.json records, per source and per image, which classes are exhaustively annotated ("full") and which only contribute positives ("partial"). In the classification loss the negative (background) term is applied only to an image's full classes; positives always count (loss * max(full_mask, target > 0)). 9,351 images have per-image overrides, e.g. 58% of one large source labels walls but not doors/windows on many sheets.
  • Label-set-aware mosaic. Mosaic augmentation only combines images with the same label set, so a fully labelled tile is never pasted next to a partially labelled one.
  • Label-set-aware validation. Predictions and ground truth are filtered to each image's full classes before scoring, which is also how all numbers on this card were computed.
  • Leak-free splits. Images are grouped by perceptual hash of the drawing (very large near-duplicate groups stay in train), with per-source validation/test budgets.
  • Rare-class repeat sampling (LVIS-style repeat factor, threshold 0.08, capped at 4x): training images with escalators are repeated 4x, exit markers about 3x, open passages about 2x (67,170 training entries in total).

Hyper-parameters (data/train_args.yaml): COCO-pretrained YOLO11s-seg, 40 epochs, SGD (lr 0.01 -> 0.0005 cosine, momentum 0.937, weight decay 5e-4, 2 warm-up epochs), batch 20 (10 per GPU), 1024 px, mosaic (closed for the last 8 epochs), scale 0.5, translate 0.1, horizontal and vertical flips, HSV jitter, AMP, mask ratio 4, max 800 detections. Trained on two RTX 3090s with Windows DDP (gloo, fp16 gradient compression) in 11.4 hours; the final epoch had the best validation fitness and is the released checkpoint. Training code: code/raqeeb_trainer.py, code/train_raqeeb.py, dataset builder code/build_dataset.py (paths are those of the original machine).

training curves

Operating thresholds

thresholds.json: per-class confidence at the class's maximum mask F1 on the validation split, clipped to [0.05, 0.9].

wall door window open_passage column stair elevator escalator exit_marker gate space_region
0.28 0.30 0.35 0.18 0.09 0.41 0.71 0.68* 0.33 0.50* 0.38

* unvalidated (fewer than 30 validation instances); detections of these classes always go to human review.

Files

path contents
raqeeb-cctv-v1.0.0.pt released weights (Ultralytics checkpoint, loads with YOLO(...), no custom code needed)
thresholds.json validation-selected per-class operating thresholds
config.json version, classes, recommended inference settings
checkpoints/ training checkpoints every 5 epochs (epoch0 ... epoch35) and last.pt, with optimizer state
metrics/ validation and test metrics (JSON + text), comparison with the previous model, training log (training_results.csv), gallery and AutoCAD checks
images/ before/after examples, AutoCAD verification images, training curves
examples/autocad/ the AutoCAD check drawing with RAQEEB layers (DXF) and the detections per floor (JSON)
code/ inference, annotation, AutoCAD, evaluation and training code
data/ label-set definitions, dataset statistics, training arguments

Licence and attribution

The weights are released under CC BY-NC 4.0 because they are trained on CubiCasa5K and FloorPlanCAD (both CC BY-NC 4.0). The Roboflow Universe datasets above are CC BY 4.0 (one is Public Domain); thanks to their authors. The AutoCAD check uses the teaching-building drawing from Men, Y. (2021), 3D indoor navigation network model, figshare, doi:10.6084/m9.figshare.15170475.v1 (CC BY 4.0). The model runs on Ultralytics YOLO, which is AGPL-3.0 licensed; using the Ultralytics library is subject to its licence.

@misc{raqeeb_cctv_2026,
  title  = {RAQEEB-CCTV: floor-plan element segmentation for CCTV planning},
  author = {{RAQEEB-CCTV contributors}},
  year   = {2026},
  version = {1.0.0},
  url    = {https://huggingface.co/swedishtech/RAQEEB-CCTV}
}
@inproceedings{kalervo2019cubicasa5k,
  title     = {CubiCasa5K: A Dataset and an Improved Multi-Task Model for Floorplan Image Analysis},
  author    = {Kalervo, Ahti and Ylioinas, Juha and H{\"a}iki{\"o}, Markus and Karhu, Antti and Kannala, Juho},
  booktitle = {Scandinavian Conference on Image Analysis (SCIA)},
  year      = {2019}
}
@inproceedings{fan2021floorplancad,
  title     = {FloorPlanCAD: A Large-Scale CAD Drawing Dataset for Panoptic Symbol Spotting},
  author    = {Fan, Zhiwen and Zhu, Lingjie and Li, Honghua and Chen, Xiaohao and Zhu, Siyu and Tan, Ping},
  booktitle = {IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2021}
}
@software{yolo11_ultralytics,
  title   = {Ultralytics YOLO11},
  author  = {Jocher, Glenn and Qiu, Jing},
  year    = {2024},
  url     = {https://github.com/ultralytics/ultralytics}
}

Version history

  • 1.0.0 (2026-11-14): first release. 62,205 training images from 16 sources, partial-label training, validation-selected thresholds, AutoCAD tool.
Downloads last month
322
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Evaluation results