Is AI superior to multimodal 3D sensor technology for transparent objects?

Christina Junger, Benjamin Simon, Gunther Notni

Abstract

Transparent objects challenge 3D perception in robotics, especially in navigation and human-robot collaboration. Conventional 3D sensors in the visible or near-infrared spectrum often fail to detect transparent materials due to their optical properties. Collecting real-world datasets for deep learning is difficult and time-consuming because ground truth acquisition requires complex preparation. Multimodal 3D sensors like thermal 3D cameras can automate dataset creation but are costly and need restrictive safety setups. Combining standard 3D sensors or RGB cameras with zero-shot deep learning models offers a promising alternative, enabling recognition of unseen transparent objects without task-specific training. However, the accuracy and feasibility of such zero-shot methods for transparent object perception remain underexplored. This paper presents an initial investigation into their potential and limitations.

Download

IMEKO-TC2-2025-014.pdf

DOI

10.21014/tc2-2025.014

IMEKO TC

TC2 - Photonics

Back to the proceedings