Tiny ImageNet 200-Class Image Classification Dataset

Tiny ImageNet contains 100,000 images across 200 classes, with 500 training, 50 validation, and 50 test images per class. The images are downsized to 64x64 colored images, and class labels are in English. It was created by zh-plus and last updated in July 2022.

Use Cases

• Train a convolutional neural network on 100,000 64x64 images to perform 200-class image classification.
• Benchmark model performance using the 50 validation images per class for hyperparameter tuning.
• Test model generalization on the held-out 50 test images per class.
• Use the 500 training images per class to study data augmentation techniques for small image resolution.

Strengths

• 100,000 total images provides a substantial sample for model training.
• 200 distinct classes offer a multi-class classification challenge.
• Structured split of 500 training, 50 validation, and 50 test images per class ensures consistent evaluation.
• Standardized 64x64 image resolution simplifies input pipeline setup.

Limitations

• Images are downsized to a low 64x64 resolution, limiting detail for fine-grained classification.
• The dataset is derived from a larger source, potentially inheriting any biases or label noise from the original ImageNet curation.
• No information is provided on the geographic or demographic distribution of image subjects.

Provenance

Source

Extended ImageNet 1K

Collection Method

Crowdsourced annotations, images downsized from a larger source dataset.

Freshness

Last updated 2022-07-12.

Geography

Region is listed as 'us', but specific spatial coverage is unknown.