OHO: A Multi-Modal, Multi-Purpose Dataset for Human-Robot Object Hand-Over.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Safe human-robot collaboration requires distinguishing hands from objects during object handover.
• Current methods lack robust solutions for real-world handover scenarios.

Purpose of the Study:

• Introduce the Object Hand-Over (OHO) dataset for developing machine learning models.
• Enable robust hand and object distinction for safety-critical robotic tasks.

Main Methods:

• Collected a dataset of objects held by hands using color, depth, and thermal imaging.
• Developed automated label generation for point clouds and image data for instance segmentation.
• Trained and evaluated an instance segmentation model for per-pixel hand-object distinction.

Main Results:

• The OHO dataset supports instance segmentation, 3D pose estimation, and shape estimation.
• Automated labels proved suitable for training effective instance segmentation models.
• Baseline experiments demonstrated successful hand-object distinction.

Conclusions:

• The OHO dataset is a valuable resource for advancing research in human-robot interaction and safety.
• The developed instance segmentation pipeline shows promise for real-world robotic applications.