Sai150/roof-top-dataset

---
license: mit
task_categories:
- image-segmentation
- object-detection
tags:
- roof
- satellite-imagery
- building
- segmentation
- autocad
- dxf
pretty_name: Roof Top Segmentation Dataset
size_categories:
- n<1K
---

# 🏠 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)
```json
{
"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)
```json
{
"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

```python
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

```python
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