AC-MVSNet: An efficient multi-view 3D reconstruction network integrating novel multi-scale feature extraction and edge enhancement

Multi-view stereo (MVS) reconstruction is a long-term research hotspot in computer vision. This paper presents a novel method, AC-MVSNet, aimed at tackling the existing challenges in multi-view stereo reconstruction, including the arduous task of processing high-resolution images, the substantial GPU memory consumption, and the problem of incomplete reconstruction. It features a new multi-scale feature extractor named CADS-Msfe and a novel depth-map optimization network with boundary enhancement, E-Refinement. Rich and precise feature information is extracted by CADS-Msfe. Subsequently, These features are inputted into the PatchMatch network to generate multi-scale depth maps. Finally, by taking advantage of the boundary enhancement effect of the E-Refinement network, the final depth maps with precise boundary information are obtained. We evaluated our proposed method on the Technical University of Denmark (DTU) and the Tanks and Temples Benchmark datasets. The results on the DTU indicate that the method in this paper enhances PatchMatchNet's completeness by 5.1 %, accuracy by 14.1 %, and overall quality by 10.5 %. It also outperforms other state-of-the-art (SOTA) methods in terms of completeness and overall quality.