Single-shot fringe projection profilometry based on multi-task learning: efficient depth reconstruction without explicit system calibrations

Fringe projection profilometry (FPP) typically captures multiple fringe patterns projected onto an object's surface to reconstruct one frame of its 3D profile. Efficiency can be greatly improved by using only a single projected fringe, which is important for real-time applications. However, it is generally believed that there is insufficient information for reliable depth reconstruction from a single-shot fringe image. In this study, we propose a multi-task learning approach to improve the accuracy and robustness of depth reconstruction from single-shot FPP, eliminating the need for tedious explicit imaging system calibration. The proposed approach was extensively validated on both synthetic and real-world datasets, and compared with other state-of-the-art deep learning methods. Experimental results demonstrate that the proposed multi-task learning method for single-shot calibrationless FPP overcomes the limitations of traditional FPP and outperforms previous deep learning methods.