mudasir13cs/floorcad-yolov8n-seg

FloorCAD Seg — Architectural Element Segmenter (YOLOv8n)

Hub: mudasir13cs/floorcad-yolov8n-seg

YOLOv8n instance segmentation model for architectural CAD floor plans. It predicts both boxes and masks for walls, openings, stairs, and interior symbols.

Companion box-only model: FloorCAD Detect (mudasir13cs/floorcad-yolov8n-detect).

Related VLM work (image → structured JSON): mudasir13cs/qwen25-vl-3b-floorplan-sft.

Example input

Original illustration (not from the training set). Use any CAD-style floor-plan raster at inference.

A second residential-style demo is in examples/demo_residential_floorplan.png.

Data source

Trained on a YOLO segmentation conversion of FloorPlanCAD-style CAD floor plans (35 symbol classes, polygon masks).

Resource	Link
Dataset paper	FloorPlanCAD: A Large-Scale CAD Drawing Dataset for Panoptic Symbol Spotting (ICCV 2021)
Project page	floorplancad.github.io
HF mirror (vector CAD / FiftyOne)	Voxel51/FloorPlanCAD
Base segmenter	Ultralytics YOLOv8n-seg (yolov8n-seg.pt)

FloorPlanCAD is a large-scale vector CAD collection with panoptic symbol labels (countable “things” and uncountable “stuff”). This checkpoint is a raster + polygon-mask YOLO conversion of that label family. Training images are not redistributed in this repo.

Classes (35)

Same taxonomy as FloorCAD Detect: doors/windows/walls, circulation (stair, elevator, escalator), wet-area fixtures, and furniture. Unnamed leftover ids: class_31, class_32, class_34, class_35.

Full map: dataset.yaml.

Load & run

from ultralytics import YOLO

model = YOLO("mudasir13cs/floorcad-yolov8n-seg")  # or local best.pt / floorcad-yolov8n-seg.pt
results = model.predict("demo_cad_floorplan.png", conf=0.25, imgsz=640)
results[0].show()

If Hub loading needs a filename:

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

ckpt = hf_hub_download("mudasir13cs/floorcad-yolov8n-seg", "floorcad-yolov8n-seg.pt")
model = YOLO(ckpt)

Training details

Architecture	YOLOv8n-seg
Image size	640
Epochs	100 (patience 20)
Batch	4
Optimizer	Ultralytics auto · AMP
Seed	0
Train images	3,715
Hardware	NVIDIA GPU (device=0)

Validation (final epoch)

	mAP50	mAP50-95
Boxes (B)	0.399	0.284
Masks (M)	0.187	0.082

Mask scores are lower than the detect-only model — CAD linework is thin and class-imbalanced; use FloorCAD Detect when you only need boxes.

Intended use

Research and prototyping: instance masks on CAD rasters, area takeoff helpers, overlay visualization. Not validated for permitting or construction sign-off.

YOLOv8 weights are AGPL-3.0. Respect FloorPlanCAD paper/project terms if you redistribute source drawings.

Citation

@InProceedings{Fan_2021_ICCV,
  author    = {Fan, Zhiwen and Zhu, Lingjie and Li, Honghua and Zhu, Siyu and Tan, Ping},
  title     = {FloorPlanCAD: A Large-Scale CAD Drawing Dataset for Panoptic Symbol Spotting},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month     = {October},
  year      = {2021},
  pages     = {10128-10137}
}

@software{ultralytics_yolov8,
  title  = {Ultralytics YOLOv8},
  author = {Jocher, Glenn and Chaurasia, Ayush and Qiu, Jing},
  url    = {https://github.com/ultralytics/ultralytics},
  license = {AGPL-3.0},
  year   = {2023}
}

Author / contact

Mudasir — Sr. AI Engineer at ECODA (에코다), building multimodal AI for architecture and building-performance workflows. MS AI Convergence, 숭실대학교 — Soongsil University, Seoul. More credentials, publications, and projects: mudasir13cs.github.io

• Hugging Face: @mudasir13cs
• GitHub: @mudasir13cs
• Email: mudasir13cs@gmail.com