ABSTRACT-50S 和 PASCAL-50S
收藏资源简介:
本研究介绍了两个新的图像描述数据集:ABSTRACT-50S 和 PASCAL-50S,由弗吉尼亚理工大学微软研究院创建。这两个数据集通过亚马逊Mechanical Turk平台收集,每张图片包含50条描述,远超现有数据集的描述数量。ABSTRACT-50S基于抽象场景数据集,包含卡通风格的抽象图像,而PASCAL-50S则基于UIUC Pascal Sentence Dataset,包含从Flickr收集的真实图像。数据集的创建过程涉及精心设计的界面和描述收集标准,确保描述的质量和多样性。这些数据集旨在推动视觉与语言领域的交叉研究,特别是在图像描述生成和理解方面,为构建更智能的交互系统提供支持。
This study introduces two novel image captioning datasets: ABSTRACT-50S and PASCAL-50S, developed by Microsoft Research at Virginia Tech. Both datasets are collected through the Amazon Mechanical Turk platform, with each image paired with 50 captions—far exceeding the caption count per image of existing datasets. ABSTRACT-50S is built upon the Abstract Scene Dataset, comprising abstract cartoon-style images, while PASCAL-50S is based on the UIUC Pascal Sentence Dataset, containing real-world images sourced from Flickr. The dataset creation process entails carefully designed interfaces and caption collection standards, which guarantee the quality and diversity of the collected captions. These datasets are intended to promote interdisciplinary research in the vision-and-language domain, especially in the areas of image captioning generation and understanding, and provide support for the development of more intelligent interactive systems.




