birder-project/TreeOfLife-10M-EOL-NaturalImages

TreeOfLife-10M-EOL-NaturalImages是一个从TreeOfLife-10M-WEBP数据集精选出来的自然生物图像数据集，经过系统的清洗过程去除了非自然内容，同时保留了高质量的生物样本。这个数据集通过多阶段的清洗流程进一步优化，包括初步清洗去除损坏或无效的图像、去除重复图像、过滤非自然图像以及基于美观分的过滤。最终得到约560万张高质量的自然图像，非常适合用于自监督学习、自然图像分类和其他需要清洁且多样化的自然世界表示的计算机视觉任务。数据集还包括预先计算的层次聚类分配和簇中心，以便于自定义采样和分析。

TreeOfLife-10M-EOL-NaturalImages is a curated subset of the TreeOfLife-10M-WEBP dataset, consisting exclusively of natural biological imagery. It has been systematically cleaned to remove non-natural content while preserving high-quality biological specimens. This dataset, with approximately 5.6 million natural images, is ideal for self-supervised learning, natural image classification, and other computer vision tasks that require a clean and diverse representation of the natural world. It includes pre-computed hierarchical K-Means clustering assignments and cluster centroids for custom sampling and analysis.