Deep-learning-based 3D reconstruction from single-shot structured-light image

With the continuous development of deep learning technology, single-shot 3D shape reconstruction combining structured light and deep learning has attracted widespread attention from researchers and has achieved a series of advancements. Deep learning-based structured light 3D reconstruction typically employs end-to-end neural network models to extract corresponding depth map from a single sinusoidal fringe. However, in addition to the most commonly used sinusoidal fringes, we are eager to understand the impact of other different structured-light patterns on depth information extraction. This paper evaluates the performance of three representative structured-light patterns (horizontal fringes, sinusoidal fringes and orthogonal fringes) in the two-dimensional (2D) image to depth map conversion task. Experimental results indicate that among these three types of structured light patterns, orthogonal pattern exhibit the best performance. The experimental results of this paper provide certain guidance for the selection of structured-light patterns in3Dmeasurement using fringe projection profilometry.