BigDetection

该数据集名为BigDetection，是一个大规模的目标检测基准，它融合了来自现有数据集（如LVIS、OpenImages和Objects365）的训练数据，创建了一个包含600个对象类别、超过3.4百万张训练图片以及3600万个边界框的统一数据集。该数据集旨在克服现有数据集的限制，通过在统一的标签空间下合并它们，为预训练目标检测器提供了更合适的数据集。其规模之大，拥有3.4百万张图片和3600万个边界框，主要任务是对目标检测器进行预训练。

BigDetection is a large-scale object detection benchmark that consolidates training data from existing datasets including LVIS, OpenImages, and Objects365 to build a unified dataset encompassing 600 object categories, over 3.4 million training images, and 36 million bounding boxes. This dataset is designed to address the limitations of existing datasets; by unifying these datasets under a consistent label space, it delivers a more appropriate resource for pre-training object detectors. With its substantial scale of 3.4 million images and 36 million bounding boxes, its primary purpose is to pre-train object detectors.