README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- object-detection
language:
- en
tags:
- image
- drone
- uav
- search-and-rescue
- person
- mav
pretty_name: forestpersons-v1
size_categories:
- 10K<n<100K
---
# ForestPersons Dataset

## Dataset Summary

ForestPersons is a large-scale dataset designed for missing person detection in forested environments under search-and-rescue scenarios. The dataset simulates realistic field conditions with varied poses (standing, sitting, lying) and visibility levels (20, 40, 70, 100). Images were captured using RGB sensors at ground and low-altitude perspectives.

## Supported Tasks

- Object Detection
- Search and Rescue Benchmarking
- Robust Detection under Dense Canopy Conditions

## Languages

- Visual data only (no textual data)

## Dataset Structure

### Data Fields

- **image**: RGB image (Electro-Optical modality)
- **annotations**: COCO-style bounding boxes with the following attributes:
  - Bounding box coordinates
  - Category: `person`
  - Visibility ratio (20, 40, 70, 100)
  - Pose (standing, sitting, lying)

### Data Splits

| Split        | # Images | # Annotations |
|--------------|----------|---------------|
| Train        | 67,686   | 145,816       |
| Validation   | 18,243   | 37,395        |
| Test         | 10,553   | 20,867        |

Total images: 96,482

## Dataset Creation

### Collection Process

Data was collected during controlled simulations of missing persons in forested areas, using human subjects posing realistically. All images were taken from heights of 1.5m to 2.0m, mimicking UAV perspectives, and were captured with:
- GoPro HERO9 Black
- Sony A57 with 24–70mm lens
- See3CAM industrial camera

Tripods were employed when drone use was impractical for safety reasons.

### Annotation Process

Annotations were manually created using [COCO Annotator](https://github.com/jsbroks/coco-annotator) by trained annotators.

### Note on Indexing

Please note that there is no sample with index 311 in this dataset. This index was intentionally skipped during dataset construction due to internal filtering steps. This does not affect dataset integrity or model training in any way.

## Usage Example

### (Recommended) Full Download — COCO Format Ready

```bash
# Clone the dataset repo (includes CSV + annotations.zip + dataset.py)
git lfs install
git clone https://huggingface.co/datasets/etri/ForestPersons
cd ForestPersons

# Download and extract all images (already included in the repo)
# Structure: images/{folder}/{image}.jpg

# Unzip COCO-style annotations
unzip annotations.zip 

# Resulting directory:
# ├── images/
# ├── annotations/
# │   ├── train.json
# │   ├── val.json
# │   └── test.json

```

### Visualize One Sample
```python
import requests
from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from io import BytesIO
from datasets import load_dataset

# Load the dataset from Hugging Face
dataset = load_dataset("etri/ForestPersons", split="validation")
sample = dataset[0]

# Image URL
BASE_URL = "https://huggingface.co/datasets/etri/ForestPersons/resolve/main/"
image_url = BASE_URL + sample["file_name"]

# Load the image, Please wait about 30 seconds....
response = requests.get(image_url)
if response.status_code == 200:
    image = Image.open(BytesIO(response.content))

    # Draw image and bounding box
    fig, ax = plt.subplots()
    ax.imshow(image)

    # Bounding box coordinates
    x = sample["bbox_x"]
    y = sample["bbox_y"]
    w = sample["bbox_w"]
    h = sample["bbox_h"]

    # Draw bounding box
    rect = patches.Rectangle((x, y), w, h, linewidth=2, edgecolor='red', facecolor='none')
    ax.add_patch(rect)

    # Draw label above the bounding box
    label = sample["category_name"]
    ax.text(x, y - 5, label, fontsize=10, color='white', backgroundcolor='red', verticalalignment='bottom')

    plt.axis("off")
    plt.show()

else:
    print(f"Failed to load image: {image_url}")
```

## License

The ForestPersons Dataset is licensed under the [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/).

Under this license, you may use, share, and adapt the dataset for non-commercial purposes, provided you give appropriate credit and distribute any derivatives under the same license.

For full license terms, please refer to the [LICENSE](./LICENSE) file.

If you have questions regarding the dataset or its usage, please contact:

**[email protected]**

## Additional Terms Regarding Trained Models

Any AI models, algorithms, or systems trained, fine-tuned, or developed using the ForestPersons Dataset are strictly limited to non-commercial use.


## Disclaimer

The ForestPersons Dataset is provided "as is" without any warranty of any kind, either express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, and non-infringement.

The authors and affiliated institutions shall not be held liable for any damages arising from the use of the dataset.

## Citation Information
If you are using this dataset, please cite
```bibtex
@misc{kim2025forestpersons,
  title     = {ForestPersons: A Large-Scale Dataset for Under-Canopy Missing Person Detection},
  author    = {Deokyun Kim and Jeongjun Lee and Jungwon Choi and Jonggeon Park and Giyoung Lee and Yookyung Kim and Myungseok Ki and Juho Lee and Jihun Cha},
  year      = {2025},
  note      = {Manuscript in preparation}
  url       = {https://huggingface.co/datasets/etri/ForestPersons}
}
```
Deokyun Kim and Jeongjun Lee and Jungwon Choi and Jonggeon Park contributed equally to this work.

## Acknowledgments

This work was supported by the Institute of Information & communications Technology Planning Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220021, Development of Core Technologies for Autonomous Searching Drones)