Accurate object pose estimation is crucial for various computer vision applications, such as human-robot interaction, augmented reality, embodied intelligence and autonomous driving. Existing methods often rely on predefined object models and categories, which can be impractical for real-world deployment. To address this challenges, this paper proposes a novel one-shot 6D pose estimation method based on multi-scale feature fusion and 3D Gaussian Splatting (3DGS) refinement. This method innovatively reconstructs the point cloud and 3D Gaussian representation of the object with integrating Structure from Motion (SfM) technology and multi-view information. A novel multi-scale feature matching fusion network is designed to effectively enhance the correspondence between images and point clouds. A 6D pose refinement module based on 3D Gaussian Splatting is proposed to improve the pose estimation accuracy. Extensive experiments and ablations conducted on LINEMOD dataset and OnePose-LowTexture dataset firmly establish the state-of-the-art performance of the proposed approach in the context of 6D object pose estimation.