Tiny ImageNet-C

Overview

Tiny ImageNet-C is a corruption-benchmarked variant of Tiny ImageNet created by applying the common corruptions and perturbations introduced by Hendrycks & Dietterich (2019). The corrupted dataset is intended to evaluate model robustness to naturalistic corruptions: each image is processed with a corruption type at multiple severity levels.

• Classes: 200 (same label set as Tiny ImageNet)

• Corruptions: multiple corruption types (see the original ImageNet-C paper for the canonical list)

• Severity levels: multiple levels (commonly 5 severity levels)

Data Structure

When accessing an example using ds[i] you will receive a dictionary with these keys:

Key	Type	Description
image	PIL.Image.Image	Corrupted RGB image (64×64)
label	int	Class label (0-199) or string class id depending on dataset builder
corruption_name	str	Name of the corruption applied (e.g., gaussian_noise, motion_blur)
corruption_level	int	Severity level for the corruption (typically 1-5)

Usage Example

from stable_datasets.images.tiny_imagenet_c import TinyImagenetC

# Loads the corrupted test split (download + prepare happens on first call)
ds = TinyImagenetC(split="test")

sample = ds[0]
print(sample.keys())  # {"image", "label", "corruption_name", "corruption_level"}

Related Datasets

• tiny_imagenet : The original (clean) Tiny ImageNet dataset

• ImageNet-C: larger corruption benchmark for ImageNet (Hendrycks et al.)

References

• Hendrycks, D. and Dietterich, T. (2019). Benchmarking Neural Network Robustness to Common Corruptions and Perturbations. ICLR.

• Zenodo archive: https://zenodo.org/records/2536630

Citation

@article{hendrycks2019robustness,
        title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
        author={Hendrycks, Dan and Dietterich, Thomas},
        journal={Proceedings of the International Conference on Learning Representations},
        year={2019}
}