ciFAIR is a variant of the popular CIFAR dataset, which uses a slightly modified test set avoiding near-duplicates between training and test data. It comprises RGB images of size 32x32 spanning 10 and 100 classes of everyday objects for ciFAIR-10 and ciFAIR-100, respectively.
CIFAR homepage: https://www.cs.toronto.edu/~kriz/cifar.html
CIFAR paper: https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf
ciFAIR homepage: https://cvjena.github.io/cifair/
ciFAIR Paper: https://arxiv.org/abs/1902.00423
We use the following splits of the ciFAIR-10/100 dataset for testing small-data performance:
| Split | Total Images | Images / Class |
|---|---|---|
| train | 300 / 3,000 | 30 |
| val | 200 / 2,000 | 20 |
| trainval | 500 / 5,000 | 50 |
| test | 10,000 | 1,000 / 100 |
train comprises the first 30 training images from each class and val the following 20.
trainval is a combination of both.
test is the full original ciFAIR-10 test set.
We achieved the following baseline performance using a Wide ResNet 16-8 trained on the trainval split and averaged over 10 runs.
| Dataset Variant | Accuracy |
|---|---|
| ciFAIR-10 | 58.22% |
| ciFAIR-100 | 53.42% |