SuperGlue

SuperGlue is Magic Leap’s graph-neural-network model for matching image features — the learned replacement for hand-crafted matchers in computer-vision pipelines. Given keypoints from two images (typically SuperPoint detections), SuperGlue decides which points correspond, robustly handling viewpoint change, occlusion and repetitive texture. The project has about 4,100 stars on GitHub.

Overview

Presented at CVPR 2020 as an oral, SuperGlue formulates matching as an assignment problem solved by a graph neural network with attention: points in each image exchange information contextually, and an optimal-transport layer produces the final correspondences, including explicit handling of unmatched points. This repository is the official release with pretrained weights for indoor (ScanNet) and outdoor (MegaDepth) scenes, demo scripts for pairs, sequences and live matching, and evaluation code for relative pose estimation. SuperGlue became a standard component in structure-from-motion, visual SLAM, visual localization and image-stitching pipelines, and its design influenced the wave of learned matchers that followed, including LightGlue from the same research lineage. The code is provided for research and evaluation, with commercial licensing available from Magic Leap.

In practice you run a detector such as SuperPoint first, then hand SuperGlue the keypoints and descriptors; its attentional graph network reasons about the whole constellation at once, which is why it succeeds on low-texture walls, repeated windows and wide-baseline pairs where nearest-neighbour matching collapses. The release includes weights trained on ScanNet for indoor scenes and MegaDepth for outdoor landmarks, plus pose-estimation evaluation on standard datasets, making it easy to reproduce the paper’s headline numbers before integrating it into a larger SLAM or reconstruction stack.

Key Features

• Learned feature matching via graph neural networks and optimal transport
• Pretrained indoor and outdoor weights included
• Pairs with SuperPoint for a complete learned local-feature pipeline
• Demos for image pairs, sequences and real-time matching
• Benchmark code for relative pose estimation
• Standard building block for SLAM, SfM and visual localization

💳 Pricing

The code and pretrained weights are free for research and evaluation use; commercial licensing is available from Magic Leap.

Who Is It For

SuperGlue is for computer-vision engineers and researchers building SLAM, 3D reconstruction, AR or localization systems who need reliable image correspondences in hard conditions.