OmniBenchmark

OmniBenchmark是一个专注于评估全视觉表示概念泛化能力的大型数据集，由南洋理工大学S-Lab创建。该数据集包含21个领域特定的子集，总计1,074,346张图像，覆盖了广泛的视觉领域，且各子集间无语义重叠。创建过程中，通过从Wikidata引入新概念丰富WordNet，确保了概念的全面性和多样性。OmniBenchmark旨在为全视觉表示学习提供一个高效且全面的评估基准，特别适用于研究模型在不同视觉领域的泛化能力。

OmniBenchmark is a large-scale dataset dedicated to evaluating the concept generalization capability of full visual representations, created by S-Lab at Nanyang Technological University. This dataset comprises 21 domain-specific subsets, with a total of 1,074,346 images covering a wide range of visual domains, and there is no semantic overlap between any subsets. During its construction, new concepts were introduced from Wikidata to enrich WordNet, ensuring the comprehensiveness and diversity of the covered concepts. OmniBenchmark aims to provide an efficient and comprehensive evaluation benchmark for full visual representation learning, and is particularly suitable for researching the generalization capabilities of models across different visual domains.