New Method Improves Long-Range 3D Imaging Accuracy

Long-range single-shot fringe projection profilometry (FPP), a technique for 3D imaging, faces significant challenges due to low signal-to-noise ratios and the inherent ill-posed nature of single-shot measurements. Existing learning-based methods often struggle beyond one meter, with a baseline UNet model showing considerable object mean absolute error. Through a detailed diagnostic process involving mechanistic interpretability and conformal uncertainty quantification, researchers discovered that the baseline models were not actually decoding fringe phase information. Instead, they were taking a "shortcut" by relying on object-boundary shape priors to infer depth, leading to inaccurate results. To rectify this, a new architecture named PhiCalNet was developed. This model is designed to output wrapped phase directly, which is then mapped to depth using a fixed, differentiable calibration layer. This architectural change effectively removes the possibility of the model using shape priors, leading to a 3.3x reduction in object mean absolute error and significantly improved accuracy in long-range 3D reconstruction.