TieredImageNet

TieredImageNet is a benchmark dataset for few-shot image classification that builds on the ImageNet collection. It is designed to evaluate how well models generalize to novel classes when only a small number of labeled examples are available, mirroring realistic low-data scenarios in visual recognition.

The dataset consists of 608 fine-grained categories organized into 34 WordNet-based superclasses. To promote generalization and reduce information leakage between training and evaluation, the classes are split into three non-overlapping sets by superclass: 20 train, 6 validation, and 8 test superclasses. Each split contains multiple classes drawn from the corresponding superclasses, ensuring that the semantic content of the test set differs from that seen during training.

TieredImageNet includes a large number of labeled images drawn from ImageNet, spanning hundreds of thousands of samples across its classes. This scale, combined with the hierarchical split, provides a challenging testbed for learning algorithms intended to transfer knowledge to unseen categories.

Evaluation with TieredImageNet typically follows episodic few-shot protocols. In an N-way K-shot task, the model must classify query images among N classes given K support examples per class. Standard benchmarks often use 5-way 1-shot and 5-way 5-shot settings, and performance is reported as accuracy with multiple episodes and seeds.

The dataset has become a widely used standard for benchmarking meta-learning, metric-learning, and transfer-learning approaches in generalized few-shot learning. It provides a more scalable and semantically distinct alternative to earlier benchmarks and is publicly available with official splits for research use.