<p>Existing depth map-based multi-view stereo (MVS) methods typically assume that texture features remain consistent across different viewpoints. However, factors such as lighting changes, occlusions, and weakly textured regions can lead to inconsistent texture features, posing challenges for feature extraction. As a result, relying solely on texture consistency does not always yield high-quality reconstruction results in certain scenarios. In contrast, high-level semantic concepts corresponding to the same objects remain consistent across different viewpoints, which we define as semantic consistency. Since designing and training new MVS networks from scratch is both costly and labor-intensive, we propose fine-tuning existing depth map-based MVS networks during testing phase by incorporating semantic consistency constraints to improve the reconstruction quality in regions with poor results. Considering the robust open-set detection and zero-shot segmentation capabilities of Grounded-SAM, we first use Grounded-SAM to generate semantic segmentation masks for arbitrary objects in multi-view images based on text instructions. These masks are then used to fine-tune pre-trained MVS networks via aligning them from different viewpoints to the reference viewpoint and optimizing the depth maps based on the proposed semantic consistency loss function. Our method is designed as a test-time approach that is adaptable to a wide range of depth map-based MVS networks, requiring only adjustments to a small number of depth-related parameters. Comprehensive experimental evaluation across different MVS networks and large-scale scenarios demonstrates that our method effectively enhances reconstruction quality at a lower computational cost.</p>