Research on Multi-view 3D Reconstruction Based on Deep Learning

End-to-end multi-view stereo (MVS) networks have achieved significant improvements in accuracy and efficiency over traditional geometric methods. However, deep learning-based multi-view reconstruction methods use convolution and variance metrics to aggregate information and form costumers, which makes it difficult for the network to effectively filter noise and perform accurate feature matching in the face of complex scenes. In addition, due to the lack of large-scale truth datasets for support, the generalization of the model to field scenes is also low. In this paper, we propose a multi-branch fusion self-supervised multi-view reconstruction network MFAE-MVSNet with a coarse-to-fine depth inference strategy to achieve high-resolution depth. Experiments with DTU datasets show that our model outperforms other self-supervised methods and most supervised methods.