Improving Depth Map Geometric Consistency Using an Enhanced Transformer Latent Code Model

In structured light 3D measurement technology, the core challenge is recovering the phase and solving for depth from a single-frame fringe pattern. The introduction of deep learning techniques has significantly enhanced the efficiency of this process. However, the depth maps generated by existing deep learning methods often exhibit noise and inaccuracies. Accurate recovery of the depth map is crucial for 3D reconstruction. This paper proposes a Transformer-based latent code prediction network to optimize the geometric consistency of depth maps. A Transformer encoder maps the depth map into a low-dimensional latent code space, where feature extraction and fusion are conducted, and a decoder reconstructs high-quality depth maps. To improve the geometric accuracy of the depth maps, a geometry-aware loss function is introduced. Our experiments, conducted using the Adam optimizer within the PyTorch framework, demonstrate that our method achieves superior performance, producing depth maps with smoother surfaces and more accurate edges.