FlexGlue uses an asymmetric design: it takes query points from the first image and searches for their corresponding locations across the second.
FlexGlue matches selected image points while rejecting lookalike scenes
The authors report that sparse-to-dense matching improves camera-pose estimation and 3D reconstruction, while joint disambiguation outperforms dedicated models at four times their throughput.
Big Tech
Philipp Lindenberger · Paul-Edouard Sarlin · Ondrej Miksik · Marc Pollefeys
ETH Zurich · Google · Microsoft Spatial AI Lab
Research Digest··3 min read
Lindenberger and colleagues introduce FlexGlue, a neural network that finds where selected points in one image appear in another, while also deciding whether the images depict overlapping scenes.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.