Meta-Album Birds: 49,054 Images of 315 Bird Species for Classification

49,054 images of 315 bird species, each resized to 128x128 pixels, form this image classification dataset. It was created by Gerald Piosenka, preprocessed by Dustin Carrion for the Meta-Album benchmark, and released in March 2022. The dataset includes images with natural backgrounds, which may introduce environmental bias.

Use Cases

• Training image classification models based on the 315 bird species categories.
• Benchmarking few-shot learning algorithms on a dataset with natural background variation.
• Studying dataset bias and class imbalance in visual recognition tasks based on the described background and gender ratio imbalances.

Strengths

• Contains 49,054 images, providing substantial data for model training.
• Includes 315 distinct classes (bird species), offering a multi-class classification challenge.
• Images are preprocessed to a uniform 128x128 pixel resolution.

Limitations

• The dataset is imbalanced regarding the ratio of male to female species images.
• Natural backgrounds may introduce bias, as some birds are frequently associated with specific environments like water.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Kaggle dataset 'BIRDS 400 - SPECIES IMAGE CLASSIFICATION' by Gerald Piosenka.

Collection Method

Images were collected and later preprocessed by resizing to 128x128 px using an anti-aliasing filter for Meta-Album.

Time Range

null

Freshness

Created 01 March 2022; last update date is unknown.

Geography

null