LM-Reloc: Levenberg-Marquardt Based Direct Visual Relocalization

TL;DR

LM-Reloc introduces a feature-agnostic direct visual relocalization method that leverages a Levenberg-Marquardt inspired loss and a pose estimation network, achieving superior accuracy over prior approaches.

Contribution

The paper proposes LM-Reloc, a novel direct visual relocalization approach that does not depend on feature matching, utilizing a Levenberg-Marquardt based loss and a pose network for improved robustness.

Findings

Outperforms previous state-of-the-art relocalization methods in accuracy.

Effective across different environmental conditions and large image baselines.

Comparable robustness to existing methods.

Abstract

We present LM-Reloc -- a novel approach for visual relocalization based on direct image alignment. In contrast to prior works that tackle the problem with a feature-based formulation, the proposed method does not rely on feature matching and RANSAC. Hence, the method can utilize not only corners but any region of the image with gradients. In particular, we propose a loss formulation inspired by the classical Levenberg-Marquardt algorithm to train LM-Net. The learned features significantly improve the robustness of direct image alignment, especially for relocalization across different conditions. To further improve the robustness of LM-Net against large image baselines, we propose a pose estimation network, CorrPoseNet, which regresses the relative pose to bootstrap the direct image alignment. Evaluations on the CARLA and Oxford RobotCar relocalization tracking benchmark show that our approach delivers more accurate results than previous state-of-the-art methods while being comparable in terms of robustness.