An efficient pretrain-finetune model for structured light 3D reconstruction

Structured light 3D reconstruction is a widely used technique for detecting surface defects in industry parts. With the development of deep learning, combining structured light 3D reconstruction with deep learning has become a research hotspot. However, directly applying a network trained on a single dataset to other datasets may not achieve the expected results. Therefore, we propose a transfer learning approach that enables neural networks to predict depth maps from gratings and be quickly applied to other datasets. The experimental results showed that the pre-trained model outperforms the original model in terms of reconstruction accuracy and processing speed, highlighting the potential of deep learning in enhancing structured light-based 3D reconstruction systems.