Touch3D
收藏资源简介:
Touch3D是由上海交通大学与蚂蚁集团联合构建的大规模三维视觉-触觉数据集,涵盖505个真实物体与超过2万次触觉接触样本。该数据集通过高精度三维扫描与GelSight Mini触觉传感器采集,融合了物体几何形态、材质纹理及多场景视觉数据,数据总量约700小时标注工时,覆盖餐饮、穿戴、工具等六类生活场景。其创建过程采用可移动采集系统与人工标注流程,确保了像素级触觉-视觉对齐质量,主要应用于机器人感知、跨模态对齐与物理交互理解,旨在解决触觉信号局部接触与全局场景对齐的学术空白。
Touch3D is a large-scale 3D vision-touch dataset jointly developed by Shanghai Jiao Tong University and Ant Group. It covers 505 real-world objects and over 20,000 tactile contact samples. Collected via high-precision 3D scanning and GelSight Mini tactile sensors, this dataset integrates object geometric shapes, material textures, and multi-scenario visual data. With a total of approximately 700 hours of annotation work, it encompasses six daily life scenarios including catering, wearables, and tools. Its construction adopts a mobile acquisition system and manual annotation workflow, ensuring pixel-level tactile-visual alignment quality. Primarily applied to robot perception, cross-modal alignment and physical interaction understanding, this dataset aims to address the academic gap in the alignment between local tactile contact signals and global scene contexts.

- 1Tac-DINO: Learning Vision-Tactile Features with Patch Alignment上海交通大学; 蚂蚁集团; 上海创新研究院 · 2026年



