Meta-Album Dogs: 120 Dog Breeds in 20,480 Images for Fine-Grained Categorization

20,480 preprocessed images of dogs across 120 distinct breeds, originally sourced from the Stanford Dogs Dataset. Researchers from Stanford University created the original dataset for the task of fine-grained image categorization, with images and annotations derived from ImageNet. The Meta-Album version, created by Dustin Carrion in March 2022, provides squared and resized 128x128 pixel images.

Use Cases

• Benchmarking fine-grained image classification models based on the 120 dog breed classes.
• Training few-shot or meta-learning algorithms using the preprocessed, standardized 128x128 image format.
• Studying intra-class variation in visual data due to factors like pose, color, and occlusion mentioned in the description.
• Developing background-invariant object recognition models, as the description notes significant background variation from man-made environments.

Strengths

• Contains 20,480 images, providing a substantial volume of data for model training.
• Covers 120 distinct dog breeds, offering a challenging fine-grained classification task.
• Images are preprocessed to a consistent 128x128 pixel resolution, simplifying input pipeline development.

Limitations

• The number of images per class and original resolution are not balanced, which may introduce bias.
• Description metadata is limited; actual data quality, such as annotation accuracy, requires manual inspection after download.
• The dataset has a large intra-class variation and significant background variation, which may increase model training difficulty.

Provenance

Source

Stanford Dogs Dataset (http://vision.stanford.edu/aditya86/ImageNetDogs/)

Collection Method

Images and annotations were sourced from ImageNet and preprocessed by cropping to square and resizing to 128x128 pixels.

Freshness

Created on 01 March 2022; last update date is unknown.