#### NOTE
This is a **Hugging Face dataset**. For large datasets, ensure `huggingface_hub>=1.1.3` to avoid rate limits. Learn more in the <a href="https://docs.voxel51.com/integrations/huggingface.html#loading-datasets-from-the-hub" target="_blank">Hugging Face integration docs</a>.

<a href="https://huggingface.co/datasets/Voxel51/parkseg12k_train" target="_blank">![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-yellow)</a>

# **Dataset Card for ParkSeg12k: Parking Lot Segmentation Dataset**

![image/png](https://huggingface.co/datasets/Voxel51/parkseg12k_train/resolve/main/parkseg12k-mq.gif)

This is a **FiftyOne** dataset with 11,355 samples from the ParkSeg12k dataset, enhanced with NDVI calculations for parking lot segmentation.

## Installation

If you haven’t already, install FiftyOne:

```bash
pip install -U fiftyone
```

## Usage

```python
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/parkseg12k_train")

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

## Dataset Details

### Dataset Description

* **Curated by:** UTEL-UIUC (Urban Traffic & Economics Lab, University of Illinois at Urbana-Champaign)
* **Enhanced by:** Harpreet Sahota (FiftyOne conversion and NDVI calculations)
* **Language(s) (NLP):** en
* **License:** See original dataset for license information

### Dataset Sources

* **Original Dataset:** https://huggingface.co/datasets/UTEL-UIUC/parkseg12k
* **Paper:** [A Pipeline and NIR-Enhanced Dataset for Parking Lot Segmentation](https://arxiv.org/pdf/2412.13179)
* **GitHub Repository:** https://github.com/UTEL-UIUC/ParkSeg12k

## Uses

### Direct Use

- Training semantic segmentation models for parking lot detection
- Urban planning and policy analysis
- Analyzing land use patterns in US cities
- Supporting parking reform policy discussions

### Out-of-Scope Use

- This dataset is specific to US cities and may not generalize to other countries
- Not suitable for real-time parking occupancy detection (detects lot boundaries, not individual spaces)

## Dataset Structure

### FiftyOne Fields

Each sample contains:

- **filepath**: Path to RGB image (512x512 pixels, 30 cm/pixel resolution)
- **segmentation**: Binary segmentation mask for parking lots (0=background, 1=parking)
- **nir**: Near-infrared channel as heatmap (upsampled from NAIP imagery)
- **ndvi**: Normalized Difference Vegetation Index heatmap (range: -1 to 1)
- **ndvi_mean**: Mean NDVI value for the image
- **ndvi_std**: Standard deviation of NDVI values
- **ndvi_min**: Minimum NDVI value
- **ndvi_max**: Maximum NDVI value

### NDVI Calculation

NDVI was computed using: `(NIR - Red) / (NIR + Red)`

- Values near 1: Dense vegetation
- Values near 0: Bare soil/pavement
- Negative values: Water bodies

This helps identify parking lot boundaries since many are surrounded by grass/vegetation.

## Dataset Creation

### Curation Rationale

Created to automate parking lot detection for urban planning discussions around minimum parking requirements (MPRs) and land use policy.

### Source Data

#### Data Collection and Processing

- RGB imagery: Google Maps satellite tiles (30 cm/pixel)
- NIR imagery: National Agriculture Imagery Program (NAIP) - upsampled from 1m/pixel to 30cm/pixel
- Covers 45 US cities with ~35,000 annotated parking lots
- Total area: 297.7 km² with 62.5 km² of labeled parking

#### Who are the source data producers?

- Google Maps (RGB imagery)
- NAIP/USDA (NIR imagery)
- Parking Reform Network (initial annotations for 42 cities)
- OpenStreetMap (additional annotations for 3 cities)

### Annotations

#### Annotation process

Manual refinement of initial annotations in QGIS, ensuring boundaries align with pavement edges rather than property lines.

#### Who are the annotators?

Students from the Urban Traffic & Economics Lab at UIUC, supervised by the paper authors.

## Personal and Sensitive Information

Satellite imagery may incidentally capture vehicles and structures but no personally identifiable information is included.

## Bias, Risks, and Limitations

- Dataset focuses on US cities; parking lot designs may differ internationally
- NIR channel contains tiling/mosaicking artifacts from orthorectification
- Temporal misalignment possible between RGB and NIR sources
- Urban-focused; may not generalize well to rural areas

## Recommendations

- Be aware of NIR artifacts when training models
- Consider using NDVI statistics to filter samples by vegetation content
- Post-processing steps (edge simplification, building/road removal) recommended for deployment

## Citation

**BibTeX:**

```bibtex
@article{qiam2024,
  title={A Pipeline and NIR-Enhanced Dataset for Parking Lot Segmentation},
  author={Shirin Qiam and Saipraneeth Devunuri and Lewis J. Lehe},
  journal={arXiv preprint arXiv:2412.13179},
  year={2024},
  url={https://arxiv.org/pdf/2412.13179}
}
```

**APA:**
Qiam, S., Devunuri, S., & Lehe, L. J. (2024). A Pipeline and NIR-Enhanced Dataset for Parking Lot Segmentation. *arXiv preprint arXiv:2412.13179*.

## Dataset Card Contact

For questions about the original dataset: {sqiam2, sd37, lehe}@illinois.edu