BDS-TR
收藏资源简介:
BDS-TR数据集是由IEEE出版技术组的研究人员开发的一个大规模显著目标检测(SOD)数据集。该数据集在DUTS-TR的基础上进行了扩展,包含了约260,000张图像,涵盖了960多个主要类别和3000多个子类别,显著提升了数据集的多样性和规模。数据集的创建过程包括从COCO、OpenImages和VOC2012等数据集中筛选图像,并通过弱监督方法生成伪标签。BDS-TR数据集旨在解决现有SOD数据集样本数量和类别有限的问题,提升模型在广泛应用场景中的泛化能力,为未来的SOD研究提供更全面的基础数据支持。
The BDS-TR dataset is a large-scale Salient Object Detection (SOD) dataset developed by researchers from the IEEE Publication Technology Group. This dataset is expanded based on the DUTS-TR dataset, containing approximately 260,000 images covering over 960 primary categories and more than 3,000 subcategories, which substantially enhances both the diversity and scale of the dataset. The development workflow of BDS-TR involves screening images from public datasets including COCO, OpenImages, and VOC2012, and generating pseudo-labels using weakly-supervised methods. The BDS-TR dataset is designed to mitigate the issues of limited sample quantity and category coverage in existing SOD datasets, improve the generalization capability of models across a wide range of application scenarios, and provide more comprehensive foundational data support for future SOD research.




