Fashion-MNIST Corrupted Out-of-Distribution Image Set

Fashion-MNIST Corrupted provides out-of-distribution data for the Fashion-MNIST dataset, based on the MNIST-C project. It contains corrupted fashion images intended for testing model robustness.

Use Cases

• Test image classification model robustness on out-of-distribution data derived from Fashion-MNIST.
• Benchmark model performance against various corruption types applied to fashion images.
• Evaluate generalization error by comparing performance on clean Fashion-MNIST versus its corrupted counterpart.

Strengths

• Based on the established MNIST-C methodology for creating corrupted data.
• Intended as a direct drop-in replacement for the standard Fashion-MNIST dataset.
• Tags indicate a size category of 10K to 100K data points.

Limitations

• Specific row count, column details, and corruption types are not provided in the input.
• The dataset's utility is dependent on the original Fashion-MNIST dataset's scope and limitations.

Provenance

Source

Extended from the Fashion-MNIST dataset, based on the MNIST-C project by Mu et al.

Collection Method

Machine-generated corruptions applied to the base Fashion-MNIST images.

Freshness

Last updated on 2024-09-10.