Meta-Album Birds: 49,054 Images of 315 Bird Species for Few-Shot Learning

Meta-Album's BRD dataset contains 49,054 images across 315 bird species, preprocessed to a resolution of 128x128 pixels. The dataset was created by Gerald Piosenka and extended by the Meta-Album project, with a release date of March 2022. Images retain their natural backgrounds, which the description notes can introduce environmental bias.

Use Cases

• Training image classification models to identify bird species based on the 315 classes.
• Benchmarking few-shot learning algorithms using the standardized Meta-Album framework.
• Studying model bias related to natural background contexts, as mentioned in the description.
• Analyzing class imbalance effects, specifically regarding male-to-female image ratios noted in the description.

Strengths

• Contains 49,054 images, providing a substantial volume for model training.
• Covers 315 distinct bird species, offering a multi-class classification challenge.
• Images are preprocessed to a consistent 128x128 pixel resolution.
• Released under a permissive CC0 Public Domain license.

Limitations

• The description notes inherent bias due to natural backgrounds (e.g., some birds frequently found in water).
• The dataset is described as imbalanced regarding the ratio of male to female species images.
• Column-level documentation and sample data are unavailable, requiring inspection after download.

Provenance

Source

Original data sourced from 'BIRDS 400 - SPECIES IMAGE CLASSIFICATION' by Gerald Piosenka on Kaggle.

Collection Method

Images were collected and later preprocessed by the Meta-Album project, resizing them to 128x128 px.

Freshness

Created 01 March 2022; last update date is unknown.