基于相对位置编码和注意力机制的多视图三维重建方法

his paper proposes a multi-view 3D reconstruction model based on Transfomer. This model achieves excellent results in the multi-view 3D reconstruction model based on deep learning. Utilizing intra-attention and inter-attention\'s ability to capture intra-image features and inter-image features enhances the network\'s ability to capture various depth information and improves accuracy. This method first uses the feature pyramid network to extract features from various scales, which enables the network to effectively perceive detailed information. In addition, this paper refers to the relative position encoding of features by the latest swin transformer, constructs an attention function suitable for multi-view reconstruction, and then proves the feasibility of this method through experiments. Finally, a depth adjustment residual network based on ResNet was also constructed. This step alleviated the problem of incomplete loss of boundary information due to training to a certain extent, and improved the quality of the overall reconstruction.