Haystack

Haystack数据集由德国奥格斯堡大学创建，专注于场景图中的罕见谓词类评估。该数据集包含超过11,300张图像，总计约25,000个关系标注，特别强调罕见谓词类。创建过程中采用模型辅助的标注流程，高效地从大量图像中识别这些罕见谓词。Haystack数据集的应用领域主要在于提升场景图生成模型对罕见谓词的理解和处理能力，解决现有数据集在罕见谓词评估上的不足。

The Haystack dataset was developed by the University of Augsburg in Germany, focusing on the evaluation of rare predicate classes in scene graphs. It contains over 11,300 images with a total of approximately 25,000 relational annotations, with special emphasis on rare predicate classes. A model-assisted annotation workflow was employed during its creation to efficiently identify these rare predicates from a large corpus of images. The core application of the Haystack dataset is to enhance the understanding and processing capabilities of scene graph generation models for rare predicates, addressing the shortcomings of existing datasets in rare predicate evaluation.