Gyroscope-Guided Optimal Transport for SuperGlue: A Training-Free Epipolar-Arc Prior
Year: 2026
Type: preprint
Abstract: Learned matchers such as SuperGlue use image evidence only, whereas visual– inertial platforms carry a gyroscope that measures the inter-frame rotation independently of image quality. Added to a pretrained SuperGlue as a log-prior around the pure-rotation prediction, this rotation did not improve pose accuracy: translation-induced parallax forces a loose tolerance, and because the dustbin receives no prior, a penalty-only prior shifts mass towards “unmatched”. We therefore propose a training-free epipolar-arc prior (EAP). With the gyroscope rotation fixed, a robust two-point solver estimates the translation direction from the same network pass, and candidates are rewarded by their angular distance to the great-circle arc on which a correct match must lie. With hyperparameters frozen on one TUM-VI sequence, the EAP raises the area under the pose-error curve at 10 • (AUC@10 •) of a 5-point pose pipeline on 15 held-out TUM-VI and EuRoC sequences (200–600 ms) by +1.9 and +2.2 points on unmodified images and by up to +9.9 points under synthetic motion blur; the gain is positive on all 15 sequences and grows with rotation. Over all regimes it gains +4.4 and +2.7 points, versus +1.0 and +0.1 for a post-hoc arc filter and +1.5 and +0.5 for a gyroscope-free arc. Transferred unchanged to LightGlue, it gains +2.3 and +0.8 points, significantly only under motion blur and, on TUM-VI, a low keypoint budget. With a pose solver that uses the gyroscope rotation, it adds only +0.5 and +0.2 points, so the EAP mainly benefits pipelines whose pose solver ignores inertial data.
Source: Research Square
Author: Ufuk Asil
Institution: Ostim Technical University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225907701
Cites:
Cited by:
Related to: