LF-MVS: Line Features-Assisted PatchMatch Multiview Stereo

As the resolution of the acquired images increases, the performance of the PatchMatch algorithm in estimating depth information for large low-textured regions is limited. The traditional PatchMatch algorithm can not extract effective texture features for matching in low-textured regions when employing a fixed-size matching window for depth estimation. In this paper, we extract the epipolar line feature on the source image for PatchMatch before depth estimation. The optimal matching window size guided by the line feature ensures that the discriminative texture can be captured, which helps improve the robustness of the depth estimation process. In addition, the extracted line features are employed to evaluate the matching ability of source images. We utilize a probabilistic graphical model to jointly model the extracted line features with pixel visibility, and propose a novel enhanced view selection scheme. Finally, a novel matching holed window is proposed to decrease the computational cost, while maintaining the robustness. The experiments performed on the ETH3D benchmark dataset demonstrate that the proposed line features-assisted Multi-View Stereo framework (LF-MVS) can effectively improve the completeness of the reconstructed point cloud, especially in weakly textured regions.