SuperGlue is a novel neural network designed to match local features between two sets of images, aiming to improve the accuracy and efficiency of feature matching in geometric computer vision tasks such as SLAM and SfM. The method leverages a graph neural network (GNN) to predict the cost of an optimal transport problem, which is used to estimate correspondences and reject non-matching points. SuperGlue incorporates attention mechanisms to handle partial point visibility and occlusion, and it is trained end-to-end from image pairs, learning priors over geometric transformations and 3D scene regularities. Compared to traditional heuristics and other learned approaches, SuperGlue outperforms state-of-the-art methods in challenging real-world indoor and outdoor environments, achieving high precision and recall in pose estimation tasks. The method is also real-time compatible and can be integrated into modern SfM or SLAM systems. The code and trained weights are publicly available.SuperGlue is a novel neural network designed to match local features between two sets of images, aiming to improve the accuracy and efficiency of feature matching in geometric computer vision tasks such as SLAM and SfM. The method leverages a graph neural network (GNN) to predict the cost of an optimal transport problem, which is used to estimate correspondences and reject non-matching points. SuperGlue incorporates attention mechanisms to handle partial point visibility and occlusion, and it is trained end-to-end from image pairs, learning priors over geometric transformations and 3D scene regularities. Compared to traditional heuristics and other learned approaches, SuperGlue outperforms state-of-the-art methods in challenging real-world indoor and outdoor environments, achieving high precision and recall in pose estimation tasks. The method is also real-time compatible and can be integrated into modern SfM or SLAM systems. The code and trained weights are publicly available.