Dataset Card for TreeOfLife-10M-EOL-NaturalImages

Dataset Description

This is a curated version of the TreeOfLife-10M-WEBP EOL training split, filtered to contain exclusively natural biological imagery. The dataset has been systematically cleaned using the Vision Data Curation (VDC) framework to remove non-natural content while preserving high-quality biological specimens.

Dataset Summary

This version further refines the dataset from birder-project/TreeOfLife-10M-WEBP through a multi-stage curation process:

• Initial Sanitization: Corrupted or invalid images were detected and removed/repaired.
• Deduplication: Near-duplicate images were identified and removed using SSCD embeddings to prevent data redundancy and improve training efficiency.
• Filtering for Natural Images: A key curation step involved extensive example-based filtering using PE-Core embeddings to remove non-natural content such as:
  • Documents (notebook pages, text)
  • Maps and charts
  • Drawings and illustrations
  • Footprint images
  • Specimen photography (where not desired for natural imagery) This step ensures the dataset is primarily composed of real-world photographs of natural subjects.
• Aesthetic Filtering: Images with very low aesthetic scores (e.g., severely blurred or out-of-focus) were identified and removed using CLIP-derived aesthetic scores, further enhancing overall visual quality.

The result is a high-quality dataset of approximately 5.6 million natural images, ideal for self-supervised learning, natural image classification, and other computer vision tasks requiring a clean and diverse representation of the natural world. This version is provided before hierarchical sampling, allowing users to apply their own sampling strategies to achieve a desired dataset size and diversity.

To facilitate custom sampling and analysis, this dataset also includes the pre-computed hierarchical K-Means clustering assignments (hierarchical_kmeans_assignments.csv) and cluster centroids (hierarchical_kmeans_centers.csv). These files can be used with the VDC framework's sampling tools (e.g., sample_images) or custom scripts to create representative subsets based on the learned cluster structure.

For a detailed walkthrough of the entire curation process that generated this dataset, please refer to the VDC Real-World Workflow: Cleaning the Tree of Life 10M Dataset documentation.

For original dataset details, licensing information, taxonomy, and annotation processes, please refer to the original TreeOfLife-10M dataset card.

Limitations

• Taxonomic coverage: Maintains original dataset limitations regarding taxonomic coverage and class imbalance among the remaining natural images
• Image resolution: Some images were resized in the preceding WEBP conversion step, which may still affect fine-grained visual analysis of extremely high-resolution specimens
• Reduced Size: The filtering process has reduced the total number of images
• Targeted Content: This dataset is specifically curated for natural images, it is not suitable for tasks requiring documents, maps, or other non-natural visual content from the original dataset

Licensing

The TreeOfLife-10M dataset compilation is released under CC0 1.0. This designation applies to the compilation and does not override the licenses of individual images or text records. The source dataset contains assets under a variety of licenses, primarily from the Creative Commons family; EOL image licenses range from CC0 to CC BY-NC-SA.

Filtering the source dataset into this curated subset does not relicense the retained assets. Their original treeoflife_id identifiers are preserved, allowing them to be linked to the upstream metadata/licenses.csv, which records each asset's license, source, copyright owner, and license URL. Users are responsible for complying with the applicable license and attribution requirements for each asset. See the original licensing information for details.

Citation

@dataset{treeoflife_10m,
  author = {Samuel Stevens and Jiaman Wu and Matthew J Thompson and Elizabeth G Campolongo and Chan Hee Song and David Edward Carlyn and Li Dong and Wasila M Dahdul and Charles Stewart and Tanya Berger-Wolf and Wei-Lun Chao and Yu Su},
  title = {TreeOfLife-10M},
  year = {2023},
  url = {https://huggingface.co/datasets/imageomics/TreeOfLife-10M},
  doi = {10.57967/hf/1972},
  publisher = {Hugging Face}
}

@inproceedings{stevens2024bioclip,
  title = {{B}io{CLIP}: A Vision Foundation Model for the Tree of Life},
  author = {Samuel Stevens and Jiaman Wu and Matthew J Thompson and Elizabeth G Campolongo and Chan Hee Song and David Edward Carlyn and Li Dong and Wasila M Dahdul and Charles Stewart and Tanya Berger-Wolf and Wei-Lun Chao and Yu Su},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2024},
  pages = {19412-19424}
}

If you utilize the Vision Data Curation (VDC) framework, please also consider citing it.

@software{Hasson_Vision_Data_Curation,
  author = {Hasson, Ofer},
  license = {Apache-2.0},
  title = {{Vision Data Curation}},
  url = {https://gitlab.com/birder/vision-data-curation}
}

Acknowledgments

This curated dataset builds upon the exceptional work of the TreeOfLife-10M creators at the Imageomics Institute. All credit for original data collection, taxonomic labeling, and scientific contributions belongs to the original team. This curation work aims to enhance the dataset's utility for computer vision research while preserving its scientific integrity.