HumanRef
收藏资源简介:
HumanRef是一个新颖的参照数据集,由国际数字经济学院(IDEA)设计,旨在解决现实世界中对任意人物的参照任务。该数据集包含103,028条参照语句,每条语句平均指代2.2个实例,涵盖了属性、位置、交互、推理和名人识别等多个方面。数据集通过人工标注和自动标注相结合的方式构建,包含多样场景中的人物图像,旨在推动人物参照表达研究的发展。
HumanRef is a novel reference dataset designed by the International Digital Economy Academy (IDEA), which aims to address the referring task for arbitrary individuals in real-world scenarios. This dataset comprises 103,028 reference utterances, with an average of 2.2 instances referred to per utterance, covering a wide range of categories including attributes, spatial positions, interactions, reasoning, and celebrity recognition. Constructed via a hybrid workflow of manual and automatic annotation, the dataset includes human images across diverse scenarios, with the goal of advancing research on person-oriented referring expressions.




