Bird Images with Bounding Boxes and Part Annotations

Containing 11,788 bird images across 200 fine-grained species subclasses, with 5,994 images for training and 5,794 for testing. Each image includes a class label, a bounding box for the bird, part location information, and species attribute data.

Use Cases

• Train a fine-grained image classifier using the 200 class labels and 11,788 images.
• Develop object detection models for birds using the provided bounding box annotations.
• Analyze bird part locations and their relationships to species attributes.
• Benchmark multi-task learning models that jointly predict class, bounding box, and part information.

Strengths

• Dataset size of 11,788 images provides a substantial sample for model training.
• Includes 200 fine-grained bird species subclasses for detailed classification tasks.
• Provides multiple annotation types per image: class label, bounding box, part locations, and attributes.

Limitations

• Dataset is from 2011 and may not reflect current bird populations or photographic styles.
• Geographic origin of the images is not specified, which may introduce bias.
• The split between training and test sets is fixed, limiting flexibility for some validation strategies.

Provenance

Source

California Institute of Technology (Caltech).

Collection Method

Curated collection of bird images with manual annotations.

Time Range

Data proposed around 2010-2011.

Freshness

Last updated on the platform in April 2022, but the underlying data is from 2011.