Five Standard Few-Shot Image Classification Datasets

Housing five standard datasets for few-shot classification, including miniImageNet and tieredImageNet. The miniImageNet dataset has 100 classes with 600 images each, partitioned into 64, 16, and 20 classes for meta-training, validation, and testing. The tieredImageNet dataset includes 608 classes grouped into 34 super-classes, with splits of 20, 6, and 8 classes for the respective meta-learning phases.

Use Cases

• Train few-shot learning models on the 100 classes of miniImageNet, using the predefined 64-class meta-training split.
• Benchmark meta-learning algorithms on tieredImageNet's 608 classes, leveraging its hierarchical structure of 34 super-classes for task construction.
• Evaluate model generalization using the separate meta-validation (16 classes in miniImageNet) and meta-testing (20 classes in miniImageNet) splits.
• Compare model performance across different dataset scales and splits, such as miniImageNet's 600 images per class versus tieredImageNet's class hierarchy.

Strengths

• The miniImageNet dataset provides a standardized benchmark with 100 classes and 600 images per class.
• Datasets include predefined splits for meta-training, meta-validation, and meta-testing, facilitating reproducible evaluation.
• The tieredImageNet dataset offers a larger scale with 608 classes organized into a hierarchical structure of 34 super-classes.

Limitations

• Specific row counts, column details, and file formats for the datasets are not provided in the input.
• The description is incomplete, referencing a full description on an external page, which may hinder immediate understanding.
• The input lacks information on image resolution, licensing, and the specifics of the three other datasets in the collection.

Provenance

Source

Built upon the ImageNet dataset.

Collection Method

Subsets of ImageNet curated for few-shot classification tasks.

Time Range

null

Freshness

null

Geography

null