Meta-Album Cars: 7,840 Images of 196 Car Models for Fine-Grained Classification

A collection of 7,840 square images depicting 196 distinct car models, preprocessed from the Stanford Cars Dataset. The images were resized to 128x128 pixels using an anti-aliasing filter after being made square by duplicating border pixels to avoid information loss. The dataset was created by Philip Boser for the Meta-Album project in March 2022.

Use Cases

• Fine-grained image classification based on 196 distinct car model classes.
• Few-shot learning benchmarking based on the standardized Meta-Album format.
• Training computer vision models to recognize specific car models from images with varied backgrounds.
• Evaluating model performance on preprocessed, uniformly sized 128x128 pixel images.

Strengths

• Contains 7,840 images across 196 classes, providing a substantial number of examples for fine-grained classification.
• Images are uniformly preprocessed to a 128x128 pixel resolution, ensuring consistency for model input.
• Preprocessing method preserved car information by duplicating border pixels instead of cropping.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• The original data was collected in 2013, which may not reflect contemporary car models.
• Image resolution is relatively low (128x128 pixels), which may limit detail for some classification tasks.

Provenance

Source

Stanford Cars Dataset, collected by Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei.

Collection Method

Images were preprocessed by creating square images via border duplication and then resized.

Time Range

Original data collected in 2013.