Learning-based reconstruction of moving transparent objects
In one sentenceFringes projected onto a transparent object reflect off both its front and rear surfaces and overlap. A diffusion model separates them; SIFT and 2D motion equations then reconstruct the moving object.
The problem
Transparent moving objects have long been challenging targets in optical imaging: they have high transmittance and weak texture, and their imaging involves both refraction and reflection. Unlike conventional opaque rigid objects, a transparent object reflects the fringe patterns projected by phase-shifting profilometry (PSP) from both its front and rear surfaces, so the captured images show severe fringe superposition. When the object moves, its refraction effects are further coupled with motion-induced errors, which makes accurate reconstruction significantly harder.
How it works
The paper systematically analyzes how the imaging mechanisms of transparent objects influence geometric modeling and motion estimation, and proposes a learning-based reconstruction method for moving transparent objects that works in two stages.
- Superimposed fringes. PSP fringe patterns are reflected by both the front and rear surfaces of the moving transparent object and overlap in the captured images.
- Fringe separation. A diffusion model separates the superimposed fringe patterns.
- Reconstruction of the moving object. The moving object is then reconstructed based on the SIFT algorithm and two-dimensional motion equations.
Results
This page summarises the paper’s abstract. The full paper is available from the publisher: doi.org/10.1117/12.3111340.
Cite
@inproceedings{feng2026learning,
title={Learning-based reconstruction of moving transparent objects},
author={Feng, Hua and Lu, Lei and Su, Zhilong and Pan, Wei},
booktitle={Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series},
volume={14175},
pages={1417511},
year={2026},
doi={10.1117/12.3111340}
}