Universal Embedding Dataset (UnED)

Universal Embedding Dataset (UnED) 是一个大规模的公共基准数据集，用于评估通用图像嵌入。该数据集由捷克技术大学和Google合作创建，包含超过400万张来自8个不同领域的图像，涵盖食品、汽车、在线产品、服装、自然世界、艺术品、地标和零售产品等多个类别。UnED的创建过程涉及精心组合现有的特定领域数据集，形成一个统一格式，具有标准分割和度量。该数据集旨在解决通用图像嵌入的问题，即训练一个能够在多个领域中使用的单一模型，以满足现代通用视觉搜索系统的需求。

Universal Embedding Dataset (UnED) is a large-scale public benchmark dataset for evaluating general-purpose image embeddings. It was co-developed by the Czech Technical University and Google, and contains over 4 million images across 8 distinct domains, covering categories including food, automobiles, online products, apparel, the natural world, artwork, landmarks, and retail products. The construction of UnED involves carefully integrating existing domain-specific datasets into a unified format with standardized data splits and evaluation metrics. This dataset aims to address the core challenge of general-purpose image embedding: training a single model that can be utilized across multiple domains to meet the requirements of modern general visual search systems.