An Efficient Multi-view Stereo Reconstruction Method Based On MA-MVSNet

Abstract

To address the problems of multi-view stereo reconstruction, such as low reconstruction completeness and poor reconstruction accuracy, we propose a novel multi-view reconstruction network named MA-MVSNet based on the improved attention mechanism. This network takes the basic MVSNet as the backbone and introduces Local-grouped Self-attention (LGSA) and Global Adaptive Average-pooling Attention (GAAA) into the reconstruction framework to make the network have both long-range dependence and local receptive field, which solves the problem that the existing convolutional neural network-based methods can not efficiently model the global contextual information of images and improves the reconstruction quality. The experiment shows that the proposed network can achieve excellent performance on DTU dataset, especially in terms of reconstruction completeness. Compared with the existing benchmark network MVSNet, our network has improved reconstruction accuracy by 5% and reconstruction completeness by 50%.