Meta_Album_PLT_NET_Mini: 25 Plant Species Images for Few-Shot Learning

120,688 images across 25 plant species, sampled from the Pl@ntNet-300k dataset. The images are 128x128 pixels, sourced and verified by citizen botanists, with each image reviewed by an average of 2.03 users. The dataset was created by Felix Herron for the Meta-Album benchmark in March 2022.

Use Cases

• Benchmarking few-shot image classification models based on the 25 plant species classes.
• Training models for automated plant species recognition from citizen-science images.
• Studying the effects of label ambiguity and long-tailed distributions in image datasets.
• Developing data augmentation techniques for small, square-format plant images.

Strengths

• Contains 120,688 images, providing substantial data for model training.
• Images have been reviewed by an average of 2.03 citizen botanists, suggesting a verification mechanism.
• Standardized image size of 128x128 pixels, which simplifies preprocessing.
• Clear class distribution is provided, with a minimum of 2,914 and maximum of 9,011 images per class.

Limitations

• Description metadata is limited; actual data quality and potential label noise require manual inspection after download.
• The dataset is a mini-version with only 25 classes, which may not represent the full diversity of the original source.
• The description mentions some images are sketches or microscope slides, indicating heterogeneity in image content.

Provenance

Source

Pl@ntNet Project, via the Pl@ntNet-300k dataset.

Collection Method

Sampled from a larger citizen-science repository, with images confirmed using a weighted user reliability score.

Geography

Global, as images are sourced by citizen botanists worldwide.