MElHuseyni/building_height_estimation

Building Contour Detection and Height Estimation Problem

Dataset Summary

The building_height_estimation dataset is a collection of satellite images with annotated building footprints (polygons) and corresponding building heights. It is designed for the joint tasks of building contour detection (segmentation) and height estimation (regression) from monocular aerial images.

• Source & Owner: The dataset originates from the AlgoTester Building Contour Detection & Height Estimation Contest.
• Hugging Face Host: MElHuseyni / building_height_estimation.
• License: MIT
• Tags: building, segmentation, regression, remote sensing

Example of a dataset sample: a 512×512 satellite image with annotated building footprints (polygon overlays) and their corresponding height values.

---

Dataset Description

Each sample consists of:

• Image: RGB satellite image (512×512 px).

• Annotations: A LabelMe-like JSON containing:

  • points: Polygon vertices (x, y) marking building footprints
  • group_id: Numeric building height (meters)

The dataset supports training and evaluation for:

• Building footprint extraction
• Height estimation from monocular imagery
• Joint segmentation + regression architectures

---

Dataset Structure

/
├── images/                # Training images
│   ├── img0001.png
│   ├── ...
│
├── ground_truth_files/    # Training labels (JSON)
│   ├── img0001.json
│   ├── ...
│
└── test_images/           # Test set (no ground truth provided)
    ├── test0001.png
    ├── ...

Example Label (JSON):

{
  "shapes": [
    {
      "points": [[316, 486], [307, 510], [312, 512]],
      "group_id": 9
    },
    {
      "points": [[416, 457], [435, 446], [421, 423], [402, 434]],
      "group_id": 7
    }
  ]
}

---

Evaluation & Scoring

The official contest defines the score as:

Score = max(0, ⌊ (Precision + Recall − 4 × HeightError) × 5 × 10⁴ ⌋)

• Precision: Correct predicted building area ÷ total predicted area
• Recall: Correctly matched ground truth area ÷ total ground truth area
• HeightError: Weighted RMSE of predicted vs. true heights for matched buildings

Important constraints:

• ≤ 1,000 buildings per image
• ≤ 300 vertices per polygon
• Total vertices per JSON ≤ 5,000
• Height ∈ [0, 1000] meters
• Coordinates within [0, 512]
• No self-intersecting polygons
• Overlap between two buildings ≤ 10% of smaller area

---

Usage

from datasets import load_dataset

ds = load_dataset("MElHuseyni/building_height_estimation")

sample = ds["train"][0]
print(sample["image"])
print(sample["buildings"])

Each building entry contains a list of polygon points and a height value.

---

Limitations

• Monocular Input: Heights are inferred from a single RGB image, no stereo or LiDAR.
• Annotation Noise: Minor misalignments or errors in footprints may exist.
• Imbalance: Height distribution may be skewed (low-rise dominant).
• Test Labels: Hidden; only evaluable via AlgoTester scoring scripts.

---

Citation

If you use this dataset, please cite both the Hugging Face dataset and the original AlgoTester contest:

@misc{building_height_estimation_MElHuseyni,
  title = {building_height_estimation},
  author = {MElHuseyni},
  year = {2025},
  howpublished = {Hugging Face Dataset},
  url = {https://huggingface.co/datasets/MElHuseyni/building_height_estimation}
}

@online{algotester_building_height,
  title = {Building Contour Detection & Height Estimation Contest},
  author = {AlgoTester},
  year = {2024},
  url = {https://algotester.com/en/ContestProblem/DisplayWithFile/135254}
}

---

Acknowledgements

• Dataset originally prepared and hosted by AlgoTester for their contest.
• Curated and published on Hugging Face by MElHuseyni.
• Licensed under MIT for research and development purposes.

---