Achieving accurate and efficient object pose estimation is a highly valuable research topic today. However, static object pose estimation is no longer sufficient for most applications, and achieving accurate pose estimation in dynamic, unstructured scenes remains a significant challenge. Most existing methods are limited in real-world scenarios due to reliance on 3D models, closed-category detection, environmental noise, and densely labeled views. In addition, template-based methods require a large number of views for retrieval, and the refinement process is insufficiently sensitive to rotational errors. We propose an object pose detection method that uses sparse 2D images and a multi-scale attention mechanism to improve pose accuracy, with a pose refiner to correct rotational errors through dense correspondences. Additionally, we performed real-time 6-DoF pose estimation in a real-world environment, and the algorithm outperforms Gen6D.Compared to Gen6D, our method shows improvements of 4.63% and 1.74% on the LINEMOD dataset under ADD and Proj-2D metrics, respectively. On LINEMOD-Occ, the improvement is 0.15%. In terms of the median rotation angle, the model achieves 24.53°, which represents a reduction of 20.32° compared to Gen6D.