Deep Learning for Single-Shot Structured Light Profilometry: A Comprehensive Dataset and Performance Analysis.

Area of Science:

• * 3D Optical Metrology
• * Computer Vision
• * Machine Learning

Background:

• * Single-shot deep learning-based structured light profilometry (SS-DL-SLP) offers rapid and robust 3D measurements.
• * Acquiring large training datasets for SS-DL-SLP is practically challenging.
• * Existing methods lack comprehensive, standardized benchmarks for evaluating deep learning models.

Purpose of the Study:

• * To introduce a large-scale, physically generated dataset for SS-DL-SLP.
• * To establish a benchmark for evaluating and comparing deep learning models in 3D optical metrology.
• * To facilitate research and promote reproducibility in SS-DL-SLP.

Main Methods:

• * Construction of a dataset with over 10,000 physical data couples.
• * Utilizing 3D-printed calibration targets with random surface profiles.
• * Recording corresponding height profiles and deformed fringe patterns.

Main Results:

• * Demonstrated high accuracy of established neural networks in predicting height information from fringe patterns.
• * Validated network robustness on unique, unseen objects.
• * Achieved reliable full-field height information retrieval.

Conclusions:

• * The presented dataset serves as a valuable resource for advancing SS-DL-SLP.
• * Publicly available code and data promote further research and development in 3D optical metrology.
• * The benchmark facilitates the exploration and comparison of novel deep learning approaches.