Radial Contour DataBase (RCDB)

RCDB是由日本产业技术综合研究所和东京工业大学合作创建的数据集，专注于通过简单的物体轮廓组合来模拟视觉识别任务。数据集包含1000个类别，每个类别由一系列轮廓图像组成，这些图像通过数学公式自动生成，避免了真实图像的隐私和版权问题。RCDB的创建旨在验证FDSL方法的有效性，特别是在没有真实图像的情况下预训练视觉变换器（ViTs）。通过调整轮廓的复杂性和参数数量，RCDB展示了在提高预训练任务难度时，精细调整的准确性也随之提高。该数据集的应用领域主要集中在图像识别和机器学习模型的预训练，旨在解决传统数据集在隐私、版权和偏见方面的问题。

RCDB is a dataset jointly created by the National Institute of Advanced Industrial Science and Technology of Japan and Tokyo Institute of Technology. It focuses on simulating visual recognition tasks by combining simple object contours. The dataset contains 1000 categories, each of which includes a series of contour images automatically generated via mathematical formulas, thus avoiding the privacy and copyright issues associated with real-world images. The development of RCDB aims to validate the effectiveness of the FDSL method, particularly for pre-training Vision Transformers (ViTs) without relying on real images. By adjusting the complexity of contours and the number of parameters, RCDB demonstrates that the fine-tuning accuracy improves as the difficulty of the pre-training task increases. The dataset is primarily applied in image recognition and pre-training of machine learning models, with the objective of addressing the privacy, copyright, and bias issues present in traditional datasets.