SAFE-MVS: Multi-View Stereo Based on Self-Attention and Feature Enhancement

Deep learning-based Multi-View Stereo (MVS) stands as a significant area of focus in computer research, to reconstruct 3D point cloud models from multi-view images. However, current MVS methods have not fully capitalized on global and local information within the feature extraction network, particularly in occluded and shadowed regions, where these approaches struggle to extract adequate image features. To tackle these challenges, this paper proposes a multi-view stereo method named SAFE-MVS based on self-attention mechanisms and feature enhancement. The core idea of this method is twofold. Firstly, it integrates self-attention mechanisms into the feature pyramid network to autonomously learn features from previous stages, thus enhancing the robustness of image features. Secondly, we incorporate a feature enhancement module into the cost volume construction stage. This module expands the receptive field of convolutional layers and incorporates positional information into the extracted features, thereby improving the model's generalization capabilities in extensive scenes. Subsequently, this paper evaluates the proposed MVS method on the DTU dataset and compares its performance against existing classical MVS methods. The experimental findings demonstrate that the proposed MVS method outperforms classical MVS methods, significantly enhancing the accuracy and completeness of point cloud models. Moreover, experimental results demonstrate that the SAFE-MVS method reconstructs point cloud models with more complete and comprehensive edges compared to the CVP-MVSNet and Cas-MVSNet methods. Additionally, the SAFE-MVS method surpasses both CVP-MVSNet and Cas-MVSNet in terms of accuracy.