Read from LIT-386 — ARXIV-1911.11763.
What it replaced
The paper is precise about its target: “traditional, hand-designed heuristics”. In matching, that means the ratio test and the filtering conventions built around it — rules that decide whether a putative correspondence is real using local evidence and a threshold somebody tuned.
SuperGlue replaces them with “priors over geometric transformations and regularities of the 3D world” learned end to end from image pairs. The mechanism is worth stating because it explains why the result is more than an accuracy delta: a graph neural network predicts assignment costs, and a differentiable optimal transport step solves the assignment — so finding correspondences and rejecting unmatchable points are one operation rather than a match followed by a filter.
Attention supplies the context: an assignment is made in view of the other assignments and the 3D structure they imply, which a per-pair threshold cannot do.
The assumption it does not question
It “matches two sets of local features”. The detector runs first, upstream, untouched — and the paper positions itself as something you drop into an existing SfM or SLAM system, which is exactly the framing that keeps the preceding stage in place.
That is not a criticism of the paper, which is excellent at what it set out to do. It is the observation that makes the next one legible: having made matching learned and strong, the remaining weakness has to be somewhere else, and LIT-387 locates it in the detector.
Why it is filed with no practice
Its recommendation — learn the matcher rather than hand-designing it — was overtaken twice within eighteen months: first by removing the detector, then by removing matching as a separate stage. Filing it as a live practice would recommend a middle step of a line whose end the record now holds at SOTA-236.
But filing nothing would leave that endpoint as a conclusion without an argument. The record gains the step where this stage was shown to be learnable at all, and the constraint it left behind, which is what the next result is about.