The paper introduces SD-MVS, a novel method for 3D reconstruction that effectively addresses challenges in textureless areas. SD-MVS integrates the Segment Anything Model (SAM) to distinguish semantic instances and leverage these constraints for pixel-wise patch deformation in matching cost and propagation. The method also employs a unique refinement strategy combining spherical coordinates and gradient descent on normals, along with pixel-wise search intervals on depths, to improve the completeness of the reconstructed 3D model. Additionally, the Expectation-Maximization (EM) algorithm is used to optimize hyperparameters, reducing the dependency on empirical tuning. Evaluations on the ETH3D benchmark and the Tanks and Temples dataset demonstrate that SD-MVS achieves state-of-the-art results with reduced time consumption and memory usage. The contributions of the paper include adaptive patch deformation, multi-scale consistency, spherical gradient refinement, and EM-based hyperparameter optimization.The paper introduces SD-MVS, a novel method for 3D reconstruction that effectively addresses challenges in textureless areas. SD-MVS integrates the Segment Anything Model (SAM) to distinguish semantic instances and leverage these constraints for pixel-wise patch deformation in matching cost and propagation. The method also employs a unique refinement strategy combining spherical coordinates and gradient descent on normals, along with pixel-wise search intervals on depths, to improve the completeness of the reconstructed 3D model. Additionally, the Expectation-Maximization (EM) algorithm is used to optimize hyperparameters, reducing the dependency on empirical tuning. Evaluations on the ETH3D benchmark and the Tanks and Temples dataset demonstrate that SD-MVS achieves state-of-the-art results with reduced time consumption and memory usage. The contributions of the paper include adaptive patch deformation, multi-scale consistency, spherical gradient refinement, and EM-based hyperparameter optimization.