Keypoint Ensemble For Image Matching

Image matching aims to identify correspondences between two images, which can be used to estimate camera pose. While the ensemble of matching methods brings benefits, it simply combines the matches and makes no match between keypoints from different methods. Moreover, the feature extractions in different methods are redundant and the method with the highest time cost brings unnecessary time overhead when parallelizing. In this paper, we propose a keypoint ensemble framework, where different kinds of keypoints are combined, and the associated descriptors are extracted from a shared feature map, facilitating the matches between different methods and less time overhead of feature extraction. To reduce the time overhead of SIFT compared to other methods in ensemble, we design Neural SIFT, which employs neural networks to imitate SIFT detector but is much more efficient. The experiments demonstrate that the ensemble of Superpoint and Neural SIFT achieves state-of-the-art performance with acceptable time cost.