<p>In 6D pose estimation, object position priors initialize 3D translation and rotation predictions. However, the Gen6D method for unseen objects lacks sufficient accuracy and robustness in detecting target positions in complex scenes, often yielding imprecise final pose estimates. To address this problem, this paper introduces the Vision Transformer encoder for the first time and proposes a new pose estimation method. Using the multi-head self-attention mechanism of ViT and traditional convolution, it integrates global context and local features to enhance the similarity score and scale offset prediction accuracy, thus improving the geometric consistency and positioning accuracy of target detection. The experiments show that the improved model outperforms other methods on both GenMOP and LINEMOD datasets. By introducing the ablation comparison of three mainstream models into the original detection network, it is further proved that the proposed method can show better detection and prediction performance on all verification targets, which verifies its effectiveness and generalization.</p>