Sai150/roof-top-dataset

🏠 Roof Top Segmentation Dataset

A dataset for roof segmentation and edge detection from satellite imagery, with annotations derived from AutoCAD DXF drawings.

Dataset Columns

Each row is one roof sample:

Column	Type	Description
sample_id	string	Unique ID (e.g. roof_001)
image	Image	Satellite roof photograph
mask	Image	Binary segmentation mask (white=roof)
overlay	Image	Verification overlay with edge lines
outer_polygon	string (JSON)	Roof boundary polygon in DXF coordinates
edges	string (JSON)	Ridge/hip/valley edge lines in DXF coordinates

AutoCAD DXF source files are available in the cad/ folder.

Annotation Format

Outer Polygon (outer_polygon column)

{
  "sample_id": "roof_001",
  "image_file": "roof_001.png",
  "image_width": 905,
  "image_height": 795,
  "class_name": "roof",
  "class_id": 0,
  "points_xy": [[x1, y1], [x2, y2], ...]
}

Edges (edges column)

{
  "sample_id": "roof_001",
  "image_file": "roof_001.png",
  "image_width": 905,
  "image_height": 795,
  "edges": [
    {"type": "line", "points": [[x1, y1], [x2, y2]]}
  ]
}

Note: Coordinates are in AutoCAD DXF world units. Use the DXF IMAGE entity's insertion point and U/V pixel vectors to convert to pixel coordinates.

DXF Coordinate → Pixel Conversion

import numpy as np, ezdxf

doc = ezdxf.readfile("cad/roof_001.dxf")
for entity in doc.modelspace():
    if entity.dxftype() == "IMAGE":
        ins = entity.dxf.insert
        u = entity.dxf.u_pixel
        v = entity.dxf.v_pixel
        h = int(entity.dxf.image_size.y)
        M_inv = np.linalg.inv([[u.x, v.x], [u.y, v.y]])
        # col, row = M_inv @ [x - ins.x, y - ins.y]
        # pixel_y = (h - 1) - row

Usage

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/roof-top-dataset")
sample = ds["train"][0]

sample["image"].show()    # satellite image
sample["mask"].show()     # binary mask
sample["overlay"].show()  # overlay visualization

import json
polygon = json.loads(sample["outer_polygon"])
edges   = json.loads(sample["edges"])

License

MIT