Ot & Sien, a dataset to help the development of object detection in children's book illustrations

The original dataset is Ot & Sien Dataset (https://lab.kb.nl/dataset/ot-sien-dataset). We corrected mistakes and made it ML-ready.The purpose of this dataset is to help the development of automatic visual object detection in children's book illustrations. The properties of our dataset are summarized as: The dataset consists of illustrations rather than standard photos. 1452 images with 8241 objects (5.7 per image) are annotated including the category and bounding boxes.All images are resized to 416 x 416 with black fitting edges to adapt to the training procedure.The dataset follows a natural long-tail property, with some object categories being rare.The dataset has imbalanced categories.

本研究所用的原始数据集为Ot & Sien数据集（Ot & Sien Dataset），官方访问链接为https://lab.kb.nl/dataset/ot-sien-dataset。我们对该数据集进行了错误修正，并将其处理为可直接用于机器学习（Machine Learning）的就绪格式。本数据集旨在助力儿童绘本插画场景下自动化视觉目标检测技术的研发。
该数据集的属性总结如下：其数据载体为插画作品而非标准摄影照片；共包含1452张图像，累计标注8241个视觉目标（单张图像平均标注5.7个目标），标注信息涵盖目标类别与边界框（bounding box）；所有图像均被统一调整至416×416分辨率，并通过添加黑色适配边缘以适配模型训练流程；该数据集呈现自然长尾分布特性，部分目标类别较为稀有；数据集的类别分布存在不平衡问题。