Meta-Album AWA Extended: 37,000 Mammal Images for Zero-Shot Learning

37,318 images of 50 animal classes, collected from public sources like Flickr in 2016 and preprocessed to 128x128 pixels. The dataset was originally created by Christoph H. Lampert, Bernt Schiele, and Zeynep Akata for benchmarking transfer-learning and zero-shot learning algorithms. This extended version was curated by the Meta-Album project and released in March 2022.

Use Cases

• Benchmarking zero-shot learning algorithms based on the dataset's design for attribute-based classification.
• Training and evaluating few-shot image classification models based on the 50 distinct animal classes.
• Developing transfer-learning models for animal recognition based on the preprocessed 128x128 pixel images.
• Researching domain adaptation in computer vision based on the dataset's origin from a specific set of public image sources.

Strengths

• Contains 37,318 images across 50 animal classes, providing a substantial scale for model training.
• Images are consistently preprocessed to a uniform 128x128 pixel resolution.
• Explicitly designed for benchmarking zero-shot and transfer-learning algorithms, providing a clear evaluation purpose.

Limitations

• Image collection date is 2016, which may not reflect current visual styles or species representations.
• Description metadata is limited; actual data quality, such as annotation accuracy, requires manual inspection after download.
• The preprocessing method for small images involved duplication of rows/columns, which may introduce artifacts.

Provenance

Source

Animals with Attributes 2 (AwA2) dataset, collected from public sources like Flickr.

Collection Method

Images were collected from public sources licensed for free use, then cropped, squared, and resized to 128x128 pixels.

Time Range

Original images collected in 2016.

Freshness

Dataset was created in March 2022, but the underlying image collection is from 2016.