Abstract: 3D Gaussian Splatting has shown significant research potential in the field of 3D reconstruction. To address the limitations of existing Gaussian-based methods in directly reconstructing target objects under complex backgrounds, this paper proposes a novel two-stage framework for object reconstruction based on multi-view consistent segmentation. In the segmentation stage, multi-reference mask prompting is introduced and cross-view attention is incorporated after feature extraction to fuse foreground features from different viewpoints, thereby enhancing the consistency and accuracy of feature representations. Additionally, a multi-view consistency loss is designed to reduce segmentation jitter caused by viewpoint variations. In the reconstruction stage, a mask-guided branch is introduced to optimize the 3DGS training process using foreground masks, improving the reconstruction loss and overall accuracy. Experimental results on the MVSeg dataset demonstrate that the proposed method achieves superior performance in terms of Chamfer Distance, with improvements of 43.5%, 38.7%, and 7.1% over SA3D, SPIn-NeRF, and SAGA, respectively. The method also shows robust reconstruction accuracy in challenging scenes with complex backgrounds and ambiguous object boundaries.