Touch100k - 大规模触觉-语言-视觉多模态数据集
收藏资源简介:
Touch100k是由北京交通大学等机构创建的大型多模态数据集,包含100,147条数据,涵盖触觉、多粒度语言和视觉模态。数据集通过收集和整理来自公开触觉数据集的101,982个视觉-触觉观察,并利用GPT-4V生成多粒度文本描述,确保数据质量。Touch100k旨在解决触觉感知研究中语言模态的不足,通过提供丰富的触觉描述,支持材料属性识别和机器人抓取预测等任务,推动触觉为中心的多模态表示学习的发展。
Touch100k is a large-scale multimodal dataset developed by institutions including Beijing Jiaotong University. It contains 100,147 data entries, covering tactile, multi-granularity linguistic, and visual modalities. The dataset is curated by collecting 101,982 visual-tactile observations from public tactile datasets, and generating multi-granularity textual descriptions via GPT-4V to ensure high data quality. Touch100k aims to address the shortage of linguistic modalities in tactile perception research, to support tasks such as material property recognition and robotic grasping prediction by providing rich tactile descriptions, and to promote the development of tactile-centric multimodal representation learning.




