In this paper, we tackle the task of estimating the 3D orientation of\npreviously-unseen objects from monocular images. This task contrasts with the\none considered by most existing deep learning methods which typically assume\nthat the testing objects have been observed during training. To handle the\nunseen objects, we follow a retrieval-based strategy and prevent the network\nfrom learning object-specific features by computing multi-scale local\nsimilarities between the query image and synthetically-generated reference\nimages. We then introduce an adaptive fusion module that robustly aggregates\nthe local similarities into a global similarity score of pairwise images.\nFurthermore, we speed up the retrieval process by developing a fast retrieval\nstrategy. Our experiments on the LineMOD, LineMOD-Occluded, and T-LESS datasets\nshow that our method yields a significantly better generalization to unseen\nobjects than previous works. Our code and pre-trained models are available at\nhttps://sailor-z.github.io/projects/Unseen_Object_Pose.html.\n