Booknis

The Booknis dataset is a collection of manually annotated images featuring medieval singing representations for segmentation and object detection. These images are derived from manuscripts that date back to the period between the 12th and 15th centuries and consist of various genres, including liturgical, secular, musical, scientific, and historical works. The dataset includes five classes, namely books, altars, lecterns, sheets, and phylacteries, which serve as indicators of potential musical representation. We have meticulously annotated 341 images and 1513 annotations, which musicology experts have verified. The primary objective of Booknis is to advance research on transfer learning by providing a more challenging benchmark for few-shot object detection.

Booknis数据集是一组经人工标注的图像集合，用于图像分割与目标检测任务，其内容聚焦中世纪歌唱场景的描绘。该数据集的图像素材均源自12至15世纪的手稿，涵盖礼仪、世俗、音乐、科学及历史等多种体裁。数据集共包含五个类别，分别为书籍（books）、祭坛（altars）、读经台（lecterns）、手稿页（sheets）与经文饰带（phylacteries），上述类别均可作为潜在音乐表现形式的识别标识。研究团队已对341张图像完成精细化标注，总计生成1513个标注样本，且该标注结果已通过音乐学专家的专业核验。Booknis数据集的核心目标在于推动迁移学习领域的研究，为少样本目标检测任务提供更具挑战性的基准测试集。