基于自注意力网络的多视角立体重建

Multi-view Stereo Reconstruction Based on Self-attention Network

为了缓解多视图立体(multi-view stereo,MVS)重建模型缺乏对图像的整体理解以及深度信息的关联未有效利用,而只关注图像局部相关性的信息,本文提出一个动态尺度特征提取网络,该网络由粗到细的级联结构输出特征以及Transformer注意力机制的方法有效地建立多视图深度相关性,适用于场景深度推断及三维重建。模型在DTU(Technical University of Denmark)公开数据集上测试,与CasMVSNet相比,重建的整体性提升6.6%。实验表明,其模型在完整性和准确性指标上较其他基于深度学习方法有一定的提升,取得较好的三维重建效果。

In order to alleviate the lack of overall understanding of the image and the ineffective use of the depth information in the MVS(multi-view stereo)reconstruction model,only the local correlation information of the image is concerned.A dynamic scale feature extraction network is used to effectively establish multi-view depth correlation by coarse-to-fine cascade structure output features and Transformer attention mechanism.The model is used for scene depth inference and 3D reconstruction.The overall of the model in the DTU(Technical University of Denmark)public data set can increase by 6.6%,which is better than the CasMVSNet reconstruction models.Experiments show that the integrity and accuracy of the model are improved compared with other methods,and better 3D reconstruction results are obtained.