We first predict dense keypoints through a single view, and use these keypoints to match instances across multiple views. We then further refine the keypoints and reconstruct the 3D point cloud through multi-view feature-level fusion of these matched instances. The final pose is solved using the Umeyama algorithm.