Meta-Album AWA Mini: 2,000 Mammal Images for Few-Shot Learning

2,000 square images of 50 mammal species, preprocessed to 128x128 pixels from the original Animals with Attributes 2 dataset. The collection was created by Meta-Album in March 2022 for benchmarking transfer and zero-shot learning algorithms. Images were sourced from public platforms like Flickr in 2016, licensed for free use and redistribution.

Use Cases

• Benchmarking zero-shot learning algorithms based on the animal attributes described in the original dataset.
• Training few-shot image classifiers for mammal species using the 50 provided classes.
• Evaluating transfer learning performance from large animal datasets to specific classification tasks.
• Studying the effect of image preprocessing (cropping and resizing to 128x128) on model accuracy.

Strengths

• Contains 2,000 images across 50 distinct animal classes, providing a multi-class benchmark.
• All images are uniformly preprocessed to a square 128x128 pixel resolution, ensuring consistent input size.
• Derived from a well-known benchmark dataset (Animals with Attributes 2) designed for transfer and zero-shot learning.

Limitations

• Dataset size is relatively small (2,000 images) compared to the original source (over 37,000 images).
• Image resolution is low (128x128 pixels) due to preprocessing, which may limit detail for fine-grained classification.
• Column-level documentation for features or attributes is absent; semantics must be inferred from the original dataset.

Provenance

Source

Animals with Attributes 2 (AWA2) dataset by Christoph H. Lampert, Bernt Schiele, Zeynep Akata.

Collection Method

Images were collected from public sources like Flickr, cropped to square, and resized to 128x128 pixels.

Time Range

Original images collected circa 2016.

Freshness

Original images were collected in 2016; the Meta-Album version was created in March 2022.