Depth Estimation for Picking Transparent Objects Using a Polarization Camera

For industrial automation, robots have to robustly pick objects with a diverse range of physical properties, such as shape, weight and surface optical properties. To realize such a purpose, this research proposes a method for depth estimation of transparent objects having complex optical properties, such as refraction and reflection from a single viewpoint. While con-ventional RGB-based or depth-completion approaches struggle to provide reliable predictions of a depth image for such transparent objects, we propose a novel monocular framework that simultaneously estimates the depth and surface normals of transparent objects from a single polarization image. Our method leverages the rich cues provided by polarization and achieves a computationally efficient depth estimation that re-quires neither analytical models of light reflection nor multi-view setups. To obtain accurate ground-truth labels for a transparent object, the proposed method uses depth and normal maps generated by existing models as pseudo ground-truth, enabling effective learning without manual labels. Experimental results demonstrate that the proposed lightweight framework achieves competitive accuracy in real-world environments.