Sun et al. (2021) — ARXIV-2104.00680
Key takeaways
- The stage that is removed is named: “instead of performing image feature detection, description, and matching sequentially”, establish pixel-wise dense matches at a coarse level and refine the good ones at a fine level.
- The reason is a failure mode, not an efficiency argument. The global receptive field of a Transformer “enables our method to produce dense matches in low-texture areas, where feature detectors usually struggle to produce repeatable interest points”. A detector cannot be fixed by a better matcher downstream; where it emits nothing, there is nothing to match.
- Descriptors conditioned on both images. Self and cross attention mean a location’s representation depends on the image it is being matched to — which a detect-then-describe pipeline cannot express, since description happens per image before any pairing exists.
- Distinguished from earlier dense methods too: those search a cost volume, this uses attention.
- Outperforms the state of the art “by a large margin”, and ranks first on two public visual localization benchmarks among published methods.
Standing in the anthology
Sources SOTA-237. The middle term of the line ending at SOTA-236: LIT-386 learns the matcher and keeps the detector, this removes the detector, and LIT-385/LIT-384 remove matching as a separate stage altogether.
Read as a pattern rather than as graphics, it is the clearest instance in this record of a recurring move: a stage survives because it was separately solvable, and removing it is what the evidence eventually supports. Each step here removes the stage the previous step revealed as the limit. That is the same shape as SOTA-227’s draft-and-verify replacing hand-tuned decoding, and the reason the geometry line is worth a language-model reader’s attention at all.