Lindenberger Philipp ; Sarlin Paul-Edouard ; Pollefeys Marc

Abstract

We introduce LightGlue, a deep neural network that learns to match localfeatures across images. We revisit multiple design decisions of SuperGlue, thestate of the art in sparse matching, and derive simple but effectiveimprovements. Cumulatively, they make LightGlue more efficient - in terms ofboth memory and computation, more accurate, and much easier to train. One keyproperty is that LightGlue is adaptive to the difficulty of the problem: theinference is much faster on image pairs that are intuitively easy to match, forexample because of a larger visual overlap or limited appearance change. Thisopens up exciting prospects for deploying deep matchers in latency-sensitiveapplications like 3D reconstruction. The code and trained models are publiclyavailable at https://github.com/cvg/LightGlue.