Deep learning profilometry for single-frame absolute 3D shape reconstruction

Recovering high-precision 3D information of dynamic scenes from single-frame fringe pattern is a major challenge in the field of fringe projection profilometry (FPP). Inspired by the successful application of deep learning in the field of FPP, we achieve single-frame, high-precision 3D measurement through the combination of data-driven and physical model-based approaches. More specifically, we utilize deep learning with powerful feature extraction ability to reduce the number of fringe images required for phase demodulation to the physical limit. Then stereo phase unwrapping (SPU) approach based on geometric constraint is used to unwrap the high-frequency wrapped phases obtained from deep learning, which maximizes the efficiency of FPP without projecting additional auxiliary patterns. Experimental results demonstrate that our method can realize high-precision 3D measurement only by a single projection, overcoming the motion sensitivity problem compared to traditional methods in dynamic scenes.