Cifair: Duplicate-Free CIFAR-10 and CIFAR-100 Test Sets

Developed by cvjena and last updated in 2021, this dataset provides a duplicate-free version of the standard CIFAR-10 and CIFAR-100 test sets. It addresses the issue of near-duplicate images between training and testing splits that can lead to inflated performance metrics in computer vision research.

Use Cases

• Benchmarking image classification models using the duplicate-free test images to ensure unbiased accuracy metrics
• Analyzing data leakage by comparing error rates on the original CIFAR test set versus this filtered version
• Validating model generalization on the 10 or 100 standard CIFAR class labels without the influence of training-set overlaps

Strengths

• Eliminates duplicate images identified in the original CIFAR-10 and CIFAR-100 test sets
• Maintains exact compatibility with existing CIFAR-10 and CIFAR-100 class labels
• Provides a more rigorous evaluation metric for 32x32 pixel image classification

Limitations

• Inherits the low-resolution 32x32 pixel limitation of the original CIFAR source
• Requires users to separately source the original training data as this release only modifies the test split

Provenance

Source

cvjena (Computer Vision Group Jena)

Collection Method

Algorithmic filtering of the original CIFAR-10/100 datasets to identify and remove duplicates.

Freshness

Last updated January 2021; static release intended for benchmarking consistency.