Meta-Album DIBaS: Digital Images of 33 Bacterial Species

4,060 preprocessed microscopic images of 33 bacterial species, derived from the DIBaS dataset. The images were split and resized to 128x128 pixels and underwent blob normalization to reduce color bias. The dataset was created by Romain Mussard for the Meta-Album benchmark in March 2022.

Use Cases

• Training few-shot image classifiers based on the 33 bacterial species classes.
• Benchmarking bias-aware computer vision models based on the blob-normalized, color-reduced images.
• Developing models for automated bacterial colony identification based on microscopic imagery.

Strengths

• Contains 4,060 images, providing at least 108 samples per class for model training.
• Images are consistently preprocessed to a uniform 128x128 pixel size.
• Applies blob normalization, a specific technique mentioned to reduce color bias for a specific challenge.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Original images were heavily augmented (split and resized), which may not reflect raw data characteristics.
• The dataset is released under a non-commercial (CC BY-NC 4.0) license, restricting commercial use.

Provenance

Source

Digital Images of Bacteria Species (DIBaS) dataset, preprocessed for Meta-Album.

Collection Method

Original microscopic images were split into smaller patches and resized.