---

name: super-glue-pretrained-network description: "Use Magic Leap SuperGluePretrainedNetwork for SuperPoint plus SuperGlue image matching, pair evaluation, live demos, and Python API workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

SuperGluePretrainedNetwork

Use this repo skill when a task involves Magic Leap's SuperGluePretrainedNetwork codebase: SuperPoint feature extraction, SuperGlue feature matching, batch pair evaluation, visualization, or direct Python use of the shipped pretrained checkpoints.

This repository is a source-only research release, not a packaged training framework. It provides inference/evaluation code, pretrained SuperPoint/SuperGlue weights, sample images and pair manifests, and two top-level command workflows. It does not release SuperGlue training code, SIFT-based SuperGlue models, or homography models.

Read first

• Repository provenance: source snapshot, evidence paths, package status, and refresh rules.
• Setup and installation: dependency expectations, source-only import model, CPU/CUDA guidance, and smoke checks.
• Model overview: SuperPoint/SuperGlue roles, indoor/outdoor weights, data flow, and output conventions.
• Troubleshooting: cross-cutting install/import, weights, backend, OpenCV, license, and performance issues.
• scripts/check_superglue_environment.py: safe environment/import/weight/backend diagnostic helper that accepts an explicit --repo-root.

Route by task

User task	Read
Write Python that imports models.Matching, SuperPoint, SuperGlue, or models.utils; inspect configs, tensors, outputs, or geometry helpers; run a tiny API smoke.	programmatic-api
Run or adapt match_pairs.py behavior for batch image-pair matching, pose evaluation, .npz outputs, pair-file validation, indoor/outdoor settings, or cache/visualization flags.	pair-matching-evaluation
Run or adapt demo_superglue.py behavior for webcam/IP/video/image-directory live matching, headless remote sequence processing, keyboard controls, or rendered match visualizations.	live-demo-and-visualization

Minimal setup check

From a SuperGluePretrainedNetwork checkout or equivalent source distribution that contains models/, models/weights/, and the two top-level scripts, install the runtime dependencies documented by the release:

python -m pip install numpy torch matplotlib opencv-python

Then run the bundled checker from this skill directory:

python scripts/check_superglue_environment.py --repo-root <superglue-repo-root>

For direct imports, the runtime Python must be able to import the repository's top-level models package. In ad hoc scripts, add the repository root to PYTHONPATH or run from the checkout root before importing:

from models.matching import Matching

Use CPU for portable checks. CUDA is optional acceleration: the repo scripts automatically choose CUDA when PyTorch reports it unless --force_cpu is set.

Operating constraints

• Inputs are grayscale images or image pairs; color images are converted to grayscale before inference in the repo utilities.
• The released pretrained weights are indoor and outdoor; choose indoor for ScanNet-like indoor pairs and outdoor for wide-baseline outdoor scenes.
• Pair evaluation requires ground-truth intrinsics and relative pose in the pair manifest. Match-only manifests should not use --eval.
• Full ScanNet/YFCC/Phototourism table reproduction needs external datasets or challenge infrastructure; the bundled samples are suitable for smoke tests and workflow validation.
• Respect the repository's noncommercial research license when using code, weights, or derivatives.

When not to use this skill

• General local feature matching libraries unrelated to SuperGluePretrainedNetwork.
• Training, fine-tuning, or checkpoint conversion requests; the release does not include training code.
• Hierarchical localization workflows outside this repo's two demo/evaluation scripts unless the task explicitly asks for SuperGluePretrainedNetwork internals.