Physics-aware depth estimation network for dense, accurate, and fast metasurface-based structured light microsystem

Structured light (SL)-based three-dimensional (3-D) reconstruction has been extensively utilized across various domains, including artificial intelligence, surveying, and precision manufacturing. In recent years, metasurface-facilitated structured light microsystems exhibit better performance in FOV, high resolution,and ultracompact volume than traditional optical projectors. However, the reconstructed point cloud data is always sparse and low-accuracy due to the limited number of projected dots depend on the fabrication size of the metasurface. In this work, we proposed a physics-aware depth estimation network for accurate depth and disparity estimation. To build a high-quality meta-optics structured light virtual dataset for training the network, our approach proposes a forward imaging process that integrates accurate meta-diffractive propagation modeling. Specifically, we achieve an absolute depth error of 0.173 mm and improve the depth-estimation accuracy by 42.56% in standard deviation compared with the common SGBM algorithm, achieving high-accuracy 3-D measurement for the recovery of the fine details of complex surfaces. Such a physics-aware depth estimation network promises accurate, fast, and dense 3-D imaging results with a compact metasurface SL module.