Improving Generalization in Fringe Projection Profilometry Networks through Physics-Informed Data Augmentation

This paper presents a novel approach to enhancing the generalization capabilities of neural networks in the fringe projection profilometry (FPP) by incorporating physical knowledge through data augmentation. Current FPP datasets primarily consist of gypsum objects, leading to performance degradation when applied to metal objects. By leveraging the physical model of FPP imaging and reflectance properties, we design a data augmentation method that simulates varying material reflectance. This approach significantly improves the model’s robustness when transferring from gypsum-based training data to metal objects. Our experimental results demonstrate an 22.8% improvement in terms of prediction MAE on metal objects compared to non-augmented models.