SuperGlue, reimplemented in PyTorch — with a slime mould twist

What it does This repo trains the SuperGlue matching network — a graph neural network plus an optimal matching layer — on pairs of SIFT keypoints and descriptors. SuperGlue sits in the “middle-end” between feature detection and downstream geometry: it aggregates context, matches keypoints across two images, and filters out the ones that have no true counterpart (occlusion, detector failure). Training runs on COCO 2014, with visualizations of predicted matches colored by confidence.

The interesting bit The stated motivation isn’t just another reimplementation: the author built it to plug in a Physarum Dynamics LP solver as a replacement for SuperGlue’s Sinkhorn algorithm. Physarum dynamics is an optimization method inspired by slime mould foraging — the companion repo actually wires it in, which is the more experimental half of this project.

Key highlights

• End-to-end PyTorch training code for SuperGlue on SIFT features, not just inference
• Enforces matching constraints: one correspondence per keypoint, with explicit unmatched handling
• Match visualizations use a jet colormap keyed to predicted confidence
• Configurable epochs, training path, and visualization output via simple flags
• Companion repo (Superglue-with-Physarum-Dynamics) explores the Sinkhorn replacement

Caveats

• The README doesn’t report any training results, accuracy numbers, or comparisons to the official implementation — you’re on your own for judging model quality
• It’s a training scaffold for a research idea more than a production matcher; the Physarum angle is the actual point

Verdict Worth a look if you’re researching differentiable matching layers or curious whether Physarum dynamics can out-Sinkhorn Sinkhorn. If you just want a working matcher for an app, the official SuperGlue (or its successors) is the safer bet.