Datasets:

MarMaster
/

corruption-shot_noise

license: mit
task_categories:
  - image-classification
language:
  - en
tags:
  - imagenet
  - corruption
  - robustness
  - computer-vision
  - image-classification
size_categories:
  - 1M<n<10M

Corruption Dataset: Shot_Noise

Dataset Description

This dataset contains corrupted versions of ImageNet-1K images using shot_noise corruption. It is part of the ImageNet-C benchmark for evaluating model robustness to common image corruptions.

Dataset Structure

Train: 1,281,167 corrupted images
Validation: 50,000 corrupted images
Classes: 1000 ImageNet-1K classes
Format: Arrow (Hugging Face Datasets)

Corruption Type: Shot_Noise

Adds shot noise (Poisson noise), common in low-light photography.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("MarMaster/corruption-shot_noise")

# Access train and validation splits
train_dataset = dataset["train"]
val_dataset = dataset["validation"]

# Example usage
for example in train_dataset:
    image = example["image"]
    class_id = example["class_id"]
    filename = example["filename"]

Dataset Statistics

Total Images: 1,331,167
Train Images: 1,281,167
Validation Images: 50,000
Classes: 1000
Image Format: RGB
Average Image Size: Variable (ImageNet-1K standard)

Citation

If you use this dataset, please cite the original ImageNet-C paper:

@article{hendrycks2019benchmarking,
  title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
  author={Hendrycks, Dan and Dietterich, Tom},
  journal={Proceedings of the International Conference on Learning Representations},
  year={2019}
}

License

This dataset is released under the MIT License. The original ImageNet dataset follows its own licensing terms.

Contact

For questions or issues, please contact: marcin.osial@[your-institution].edu