AndreaPorri/Eyes-Detection

Accurate Eyes Detection Dataset

Dataset Description

This COCO dataset is designed for real-time "precise eye detection" task under various conditions. It contains highly accurate bounding box annotations, manually retraced with Roboflow to ensure maximum precision.

Dataset Splits

This dataset is intentionally provided as a single training split containing all 72,317 examples. This design choice allows researchers to:

• Create custom split ratios tailored to their specific needs.
• Ensure different random seeds don't lead to overlapping examples between studies.
• Implement cross-validation strategies more flexibly.

You are responsible for creating appropriate validation and test splits using the provided training data. We recommend using 70-80% for training, 10-15% for validation, and 10-15% for testing, depending on your requirements.

Key features

• 72,317 high-resolution images (640×640 pixels, JPG format) featuring close-up eyes, people's faces, and real-life scenes with varying complexity and number of faces.
• All bounding boxes have been completely re-annotated manually with Roboflow to ensure maximum accuracy.
• Annotations provided in ready-to-use COCO format for easy integration with most object detection frameworks.
• Single Training Split: The dataset is provided as a single training split to allow maximum flexibility in creating custom train/validation/test splits for different research needs.
• Various Image acquisition conditions:
  • Single person (1 or 2 eyes), whole people, groups, ...
  • Open and closed eyes.
  • Diversity of lighting and poses.

Repository File Structure

The repository contains the following files. The standard structure for a COCO dataset on Hugging Face is automatically created when loaded.

Eyes-Detection/
├── README.md                 # This dataset card
├── dataset_info.json         # Dataset metadata for Hugging Face
├── annotations.coco.json     # Annotations in COCO format
├── images.zip                # All images in a single compressed archive
├── README.dataset.txt        # Marginal info
└── LICENSE                   # The CC BY-NC-SA 4.0 license file

Loading the Dataset with Hugging Face 🤗

This dataset is in COCO format and can be easily loaded using the datasets library distributed by HuggingFace.

Important Notes:

• Single Split: This dataset contains only a training split (train). You will need to create your own validation and test splits programmatically.
• Images: The images are provided in a compressed ZIP file (images.zip).
• Automatic Extraction: The load_dataset function will automatically extract images from the images.zip file.

Example Code:

from datasets import load_dataset
from sklearn.model_selection import train_test_split
import numpy as np

# Load the entire dataset as the training split
dataset = load_dataset("AndreaPorri/Eyes-Detection")
full_train_dataset = dataset['train']  # This contains all 72,317 examples

# It is NECESSARY to create a three-part division (training/validation/testing, to obtain and verify results reliably) and keep the test set completely separate from other data during model development and hyperparameter tuning.
# First split: 80% train, 20% temp (for val+test)
first_split = full_train_dataset.train_test_split(test_size=0.2, seed=42)

# Second split: split temp into 50% validation, 50% test (10% each of total)
second_split = first_split['test'].train_test_split(test_size=0.5, seed=42)

final_dataset = {
    'train': first_split['train'],
    'validation': second_split['train'], 
    'test': second_split['test']
}
 

Note: When using load_dataset, the extracted images will be cached in the ~/.cache/huggingface/datasets directory, creating the standard folder structure expected by the COCO loader.

For Advanced Usage (Manual Extraction or COCO Tools):

If you prefer to handle the extraction manually or use the annotations directly with other libraries:

1. Manual Extraction:
   Download and extract images.zip to a folder. You can then use the data_dir parameter:

dataset = load_dataset("AndreaPorri/Eyes-Detection", data_dir="path/to/extracted_images")

2. Using COCO Annotations Directly:
   The annotations.coco.json file follows the standard COCO format and can be used with libraries like torchvision or pycocotools.

from torchvision.datasets import CocoDetection
import torchvision.transforms as transforms

coco_dataset = CocoDetection(
    root='path/to/extracted_images',
    annFile='path/to/annotations.coco.json',
    transform=transforms.ToTensor()
)

Preprocessing e Augmentation applied

1. Preprocessing:

   • Auto-orientation (with EXIF orientation stripping).
   • Resize to 640x640 (Stretch).

2. Data Augmentation (create 7 versions for each image):

   • Random crop: 0-30% of the image.
   • Random rotation: ±15 degrees.
   • Random shear: ±15° horizontal/vertical.
   • Brightness adjustment: ±25%.
   • Exposure adjustment: ±15%.
   • Gaussian blur: up to 1.2px.
   • Salt and pepper noise: 0.3% of pixels.
   • Cutout: 7 boxes, each with a size of 2%.

Original Datasets of Origin

This dataset was created based on:

1. EyeCon Dataset - https://app.roboflow.com/andreap/eyecon-eaux0-oykgj/1
2. Eyes Detection Dataset - https://app.roboflow.com/andreap/eyes_detection-bupne/1

Expected Uses

• Real-time eye detection.
• Part of an Eye-Tracking pipeline for computer vision applications.
• Academic research.
• Research for medical applications.

Limitations and Ethical Considerations

• The dataset contains images of faces.
• Use permitted only for non-commercial purposes.
• It is the user's responsibility to ensure ethical use in compliance with privacy laws.

License - CC BY-NC-SA 4.0

1. What you can do:

   • Share - copy and redistribute the material.
   • Adapt - remix, transform, and create derivative works.
   • Use for research, study, and non-commercial projects.

2. Mandatory conditions:

   • Attribution - You must give appropriate credit, provide a link to the license, and indicate if changes were made.
   • NonCommercial - You may not use the material for commercial purposes.
   • ShareAlike - If you remix, transform, or create derivative works, you must distribute your contributions under the same license as the original.

3. Example of attribution:
   "Accurate Eyes Detection Dataset" by AndreaPorri is licensed under CC BY-NC-SA 4.0
   Dataset source: https://huggingface.co/datasets/AndreaPorri/Eyes-Detection

Disclaimer

The user is responsible for the ethical and legal use of this dataset. The creator assumes no responsibility for any improper use.

Contacts

For questions or additional information:

• Email: [email protected]

Citation

If you use this dataset in your research, please cite it as:

@dataset{accurate_eyes_detection_2025,
  author = {Andrea Porri},
  title = {Accurate Eyes Detection Dataset},
  year = {2025},
  publisher = {Hugging Face},
  version = {2.0},
  license = {CC BY-NC-SA 4.0},
  url = {https://huggingface.co/datasets/AndreaPorri/Eyes-Detection}
}