Note

This is a Hugging Face dataset. Learn how to load datasets from the Hub in the Hugging Face integration docs.

Hugging Face

Dataset Card for FloorPlanCAD (test split)#

image

This is a FiftyOne dataset with 5308 samples.

Installation#

If you haven’t already, install FiftyOne:

pip install -U fiftyone

Usage#

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/FloorPlanCAD")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details#

Dataset Description#

FloorPlanCAD is a large-scale real-world CAD drawing dataset containing over 15,000 annotated floor plans for panoptic symbol spotting in architectural drawings. The dataset provides line-grained vector annotations for 30 object categories across residential and commercial buildings.

Key Features:

  • Format: Vector graphics (SVG) with corresponding PNG rasterizations

  • Scale: 15,663 CAD drawings (originally 10,094 in v1, updated to 15,663)

  • Categories: 30 classes total

    • 28 “thing” classes (countable instances): doors, windows, furniture, appliances, equipment

    • 2 “stuff” classes (semantic regions): wall, parking

  • Annotation Type: Line-grained primitive-level annotations with semantic and instance labels

  • Original Split: 6,382 training / 3,712 testing drawings

  • Privacy Protected: Cropped into 20m Ă— 20m blocks, 50% retention rate, sensitive text removed

  • Curated by: Zhiwen Fan†, Lingjie Zhu†, Honghua Li, Xiaohao Chen, Siyu Zhu, Ping Tan (Alibaba A.I. Labs & Simon Fraser University, †Equal contribution)

  • Funded by: Alibaba A.I. Labs

  • Language(s): Not applicable (architectural vector graphics)

  • License: Creative Commons Attribution-NonCommercial 4.0 License

  • Project shutdown notice: As of January 2023, the project was shut down and most participants left the company

Dataset Sources#

  • Repository: https://floorplancad.github.io/ (Note: Project shut down in 2022)

  • Paper: Fan et al. “FloorPlanCAD: A Large-Scale CAD Drawing Dataset for Panoptic Symbol Spotting” (ICCV 2021)

Uses#

Direct Use#

This dataset is designed for:

  • Panoptic symbol spotting: Detecting both countable object instances and semantic regions in architectural drawings

  • Instance segmentation: Identifying individual furniture, fixtures, and building elements

  • Semantic segmentation: Recognizing structural elements like walls and parking areas

  • CAD drawing analysis: Training models for automated floor plan understanding

  • Architecture/Engineering/Construction (AEC) applications: Automated 3D modeling from 2D CAD drawings

Out-of-Scope Use#

  • Commercial applications: Dataset is licensed for non-commercial use only

  • Privacy-sensitive reconstruction: The dataset is intentionally cropped and anonymized; attempting to reconstruct original complete floor plans or identify building locations violates privacy protections

  • As-is architectural design: The cropped 20m Ă— 20m blocks are not complete floor plans suitable for construction

Dataset Structure#

The converted FiftyOne dataset contains the following structure:

<Sample: {
    'id': '690a547c0420c654cb79d521',
    'media_type': 'image',
    'filepath': '../image_data/0000-0003.png',
    'tags': [],
    'metadata': <ImageMetadata: {
        'size_bytes': 7803,
        'mime_type': 'image/png',
        'width': 1000,
        'height': 1000,
        'num_channels': 4,
    }>,
    'created_at': datetime.datetime(2025, 11, 4, 19, 31, 8, 427000),
    'last_modified_at': datetime.datetime(2025, 11, 4, 19, 39, 58, 326000),
    'ground_truth': <Detections: {
        'detections': [
            <Detection: {
                'id': '690a547c0420c654cb79d520',
                'attributes': {},
                'tags': [],
                'label': 'wall',
                'bounding_box': [0.30975255, 0.0, 0.69024745, 0.7205705549999999],
                'mask': array([[255, 255, 255, ...,   0,   0,   0],
                       [255, 255, 255, ...,   0,   0,   0],
                       [255, 255, 255, ...,   0,   0,   0],
                       ...,
                       [255, 255, 255, ...,   0,   0,   0],
                       [255, 255, 255, ..., 255, 255, 255],
                       [255, 255, 255, ..., 255, 255, 255]], dtype=uint8),
                'mask_path': None,
                'confidence': None,
                'index': None,
            }>,
        ],
    }>,
}>

Object Categories (30 total):

Doors (3): single_door, double_door, sliding_door

Windows (4): window, bay_window, blind_window, opening_symbol

Stairs (1): stair

Home Appliances (3): gas_stove, refrigerator, washing_machine

Furniture (11): sofa, bed, chair, table, bedside_cupboard, tv_cabinet, half_height_cabinet, high_cabinet, wardrobe, sink, bath

Equipment (6): bath_tub, squat_toilet, urinal, toilet, elevator, escalator

Stuff Classes (2): wall, parking

Note on class distribution: Wall and parking together account for ~27% of all annotated primitives. Significant class imbalance exists across categories.

Dataset Creation#

Curation Rationale#

The FloorPlanCAD dataset was created to address critical limitations in existing symbol spotting research:

  1. Scale: Previous datasets (SESYD with 1,000 synthetic plans, FPLAN-POLY with 42 plans) were too small for deep learning

  2. Real-world diversity: Prior datasets lacked the symbol variation seen across different architectural firms and building types

  3. Vector graphics: Maintaining CAD’s native vector format (rather than rasterization) preserves accuracy and enables graph-based methods

  4. Panoptic scope: Traditional symbol spotting focused only on “thing” instances; this dataset includes “stuff” classes (walls, parking) for complete scene understanding

Source Data#

Data Collection and Processing#

Original Data Sources:

  • 100+ architectural projects from production environments

  • Multiple partner companies and institutions

  • Building types: residential towers, schools, hospitals, shopping malls, office buildings

  • Geographic diversity: Projects from various regions (layer names include Chinese characters indicating Asian sources)

Technical Processing:

  • Multi-layer SVG organization (dozens of layers per drawing)

  • Layer-by-layer annotation to reduce clutter

  • Scale handling: Entity lengths range from millimeters to tens of meters (5+ orders of magnitude)

  • Coordinate systems: Metric units (meters) for real-world measurements

Who are the source data producers?#

  • Primary producers: Architects, engineers, and CAD designers from various companies creating production floor plans

  • Data providers: Multiple partner companies and institutions in the AEC industry (anonymized for privacy)

  • Geographic origin: Multinational (layer names suggest significant Asian representation)

Annotations#

Annotation Process#

From SVG to FiftyOne Annotations:

The conversion from vector SVG to structured annotations involves several stages:

  1. SVG Primitive Parsing (using svgpathtools):

    # For each <path>, <circle>, <ellipse> element:
- Extract semantic-id (class label 1-35)
- Extract instance-id (unique instance number or -1 for stuff)
- Parse geometry:
  * Paths: start point, end point, middle point via path.point(0.5)
  * Circles: center (cx, cy), radius (r)
  * Ellipses: center, radii (rx, ry)

  2. Coordinate Transformation:

    • SVG coordinates scaled by 10x to match PNG dimensions

    • svg_x * 10 → png_x

    • Maintains accurate spatial relationships

  3. Instance Grouping:

    • Primitives grouped by (semantic_id, instance_id) tuple

    • Each unique tuple represents one object instance

    • “Stuff” classes have instance_id = -1 (no individual instances)

  4. Bounding Box Computation:

    # For each instance:
- Collect all primitive endpoints and centers
- Compute axis-aligned bounding box:
  x_min = min(all_x_coordinates)
  y_min = min(all_y_coordinates)
  width = x_max - x_min
  height = y_max - y_min
- Normalize to [0, 1] by dividing by image dimensions

  5. Segmentation Mask Rendering:

    # For each instance:
- Create blank mask (image_height Ă— image_width)
- Render each primitive with line_width=3 pixels:
  * Paths: cv2.line() or cv2.polylines()
  * Circles: cv2.circle()
  * Ellipses: cv2.ellipse()
- Crop mask to bounding box region

  6. FiftyOne Detection Object Creation:

    • Each instance becomes fo.Detection() with:

      • label: mapped class name (e.g., “wall”, “single_door”)

      • bounding_box: normalized [x, y, w, h]

      • mask: binary array (if include_masks=True)

Who are the annotators?#

  • Number: 11 specialist annotators

  • Time investment: Over 1,000 hours total annotation effort

  • Expertise: Domain specialists familiar with architectural CAD drawings

  • Quality control: Layer-by-layer annotation methodology for accuracy

  • Employer: Alibaba A.I. Labs (annotation team)

Citation#

BibTeX#

@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}
}

APA#

Fan, Z., Zhu, L., Li, H., Zhu, S., & Tan, P. (2021). FloorPlanCAD: A Large-Scale CAD Drawing Dataset for Panoptic Symbol Spotting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 10128-10137).