Mini ImageNet: 100 Classes with Original Image Resolutions

65,000 images across 100 classes from the ImageNet-1k dataset, curated by the timm library author in 2024. It provides a high-fidelity subset for classification tasks by maintaining original image dimensions instead of downsampling to low resolutions.

Use Cases

• Training classification models on the 100 provided class labels
• Few-shot learning evaluation using the 5,000-sample test split
• High-resolution feature extraction for transfer learning

Strengths

• 65,000 total samples
• Original image resolutions preserved
• Balanced test split with 50 samples per class

Limitations

• Limited to 10% of the original ImageNet-1k class variety
• Inherits taxonomic and geographic biases present in the parent ImageNet dataset

Provenance

Source

ImageNet (ILSVRC-2012)

Collection Method

Subsampled from ImageNet-1k

Freshness

Last updated November 2024.

Geography

Global