Touch-and-Go (TG); Web-Material
收藏资源简介:
该研究构建了Touch-and-Go (TG)和Web-Material两个触觉-视觉对齐数据集,旨在解决跨模态材料分割任务。Web-Material包含约1.7万张从网络采集的野外场景图像,覆盖多种材料类别,并通过大语言模型生成多样化查询词增强数据多样性。数据集通过CLIP相似度过滤误分类样本,确保材料标注准确性。其核心创新在于提出材料多样性配对策略,将触觉信号与视觉多样的同类别图像对齐,用于训练局部跨模态特征匹配模型,推动机器人触觉感知与场景理解的研究。
This study constructs two tactile-visual alignment datasets, Touch-and-Go (TG) and Web-Material, aiming to address the cross-modal material segmentation task. Web-Material contains approximately 17,000 wild scene images collected from the web, covering a variety of material categories, and enhances data diversity by generating diverse query terms through large language models. This dataset filters misclassified samples using CLIP similarity to guarantee the accuracy of material annotations. Its core innovation is the proposal of a material diversity pairing strategy, which aligns tactile signals with visually diverse images of the same material category for training local cross-modal feature matching models, thereby advancing research on robotic tactile perception and scene understanding.
- 1Seeing Through Touch: Tactile-Driven Visual Localization of Material Regions韩国科学技术院; 韩国外国语大学; 蔚山国立科学技术院 · 2026年



