FusionMVS: Leveraging Attention and Visual Compression for Multi-View Stereo

Recent advancements in multi-view stereo (MVS) have significantly improved 3D reconstruction, yet existing methods struggle with complex scenes, low-texture regions, and occlusions. These challenges arise due to the inadequate integration of multi-scale global context and cross-channel features during feature extraction, resulting in poor detail capture in low-texture areas. Furthermore, geometric consistency across multiple views is insufficiently modeled during the cost volume regularization phase, particularly in occluded regions and complex scenarios with inconsistent viewpoints. Additionally, traditional depth regression methods fail to optimize the cost volume probability distribution, causing errors to amplify during depth estimation. To address these issues, we introduce Fusion-MVS, a novel framework incorporating the Enhanced Visual Compression Block (EVCBlock) and Convolutional Block Attention Module (CBAM). EVCBlock enhances feature extraction in low-texture regions by using deep convolution, max pooling, and crosschannel fusion. CBAM strengthens global context modeling during cost volume regularization by integrating channel and spatial attention mechanisms, improving geometric consistency in complex scenes. Additionally, a fine-grained depth-guided dynamic regression strategy is proposed to optimize cost volume generation and mitigate error propagation through expectation-based depth regression. Experimental results show that FusionMVS outperforms state-of-the-art methods on the DTU dataset, achieving significant improvements in depth estimation accuracy, completeness, and overall reconstruction quality. This approach advances MVS technology for complex scenarios and provides a novel framework for high-precision 3D reconstruction.