LightGlue: Local Feature Matching at Light Speed | VitaVision
PaperICCV · 2023

LightGlue: Local Feature Matching at Light Speed

Philipp LindenbergerPaul-Edouard SarlinMarc Pollefeys
arXiv 2306.13643DOIView in graph
1page built on it
5more pages cite it
2narratives
0registry papers cite it

In the Atlas

ModelPrimary source of
LightGlue

Adaptive-depth Transformer matcher for sparse local features: stacks 9 self+cross-attention layers with rotary positional encoding and a per-token confidence head, exits early on easy image pairs, and replaces SuperGlue's Sinkhorn solver with a dual-softmax × matchability assignment head — over 2× faster than SuperGlue at equivalent or better pose-estimation accuracy.

Cited as a reference by 5 pages, grouped by domain

Features 3
Feature MatchingLoFTRSuperGlue
Representation learning 2
Attention MechanismPositional Encoding

Appears in 2 narratives

NarrativeFoundation Models for VisionNarrativeThe Geometry Pipeline Collapses

Lineage

Papers in the Atlas registry only · one dot per paper
Built upon by · 0
’04’05’06’07’08’09’10’11’12’13’14’15’16’17’18’19’20’21’22’23Distinctive Image Features from Scale-Invariant Keypoints (2004)SuperPoint: Self-Supervised Interest Point Detection and Description (2018)SuperGlue: Learning Feature Matching with Graph Neural Networks (2020)LoFTR: Detector-Free Local Feature Matching with Transformers (2021)
Built upon by · 0