Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LiteVLoc: Map-Lite Visual Localization for Image Goal Navigation

Jianhao Jiao, Jinhao He, Changkun Liu, Sebastian Aegidius, Xiangcheng Hu, Tristan Braud, Dimitrios Kanoulas

TL;DR

Lite VLoc is a hierarchical vi-sual localization framework that uses a lightweight topo-metric map to represent the environment and uses a learning-based feature matcher to establish dense correspondences between sparse keyframes and observations, and then refines poses with a geometric solver, enabling robustness to viewpoint changes.

Abstract

This paper presents LiteVLoc, a hierarchical visual localization framework that uses a lightweight topo-metric map to represent the environment. The method consists of three sequential modules that estimate camera poses in a coarse-to-fine manner. Unlike mainstream approaches relying on detailed 3D representations, LiteVLoc reduces storage overhead by leveraging learning-based feature matching and geometric solvers for metric pose estimation. A novel dataset for the map-free relocalization task is also introduced. Extensive experiments including localization and navigation in both simulated and real-world scenarios have validate the system's performance and demonstrated its precision and efficiency for large-scale deployment. Code and data will be made publicly available.

LiteVLoc: Map-Lite Visual Localization for Image Goal Navigation

TL;DR

Lite VLoc is a hierarchical vi-sual localization framework that uses a lightweight topo-metric map to represent the environment and uses a learning-based feature matcher to establish dense correspondences between sparse keyframes and observations, and then refines poses with a geometric solver, enabling robustness to viewpoint changes.

Abstract

This paper presents LiteVLoc, a hierarchical visual localization framework that uses a lightweight topo-metric map to represent the environment. The method consists of three sequential modules that estimate camera poses in a coarse-to-fine manner. Unlike mainstream approaches relying on detailed 3D representations, LiteVLoc reduces storage overhead by leveraging learning-based feature matching and geometric solvers for metric pose estimation. A novel dataset for the map-free relocalization task is also introduced. Extensive experiments including localization and navigation in both simulated and real-world scenarios have validate the system's performance and demonstrated its precision and efficiency for large-scale deployment. Code and data will be made publicly available.
Paper Structure (30 sections, 2 equations, 4 figures, 4 tables)

This paper contains 30 sections, 2 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Motivation
  3. Contributions
  4. Related Work
  5. VLoc without 3D Representation
  6. VLoc with 3D Representation
  7. Methodology
  8. Preliminary
  9. Problem Definition
  10. LiteVLoc Framework
  11. Map Construction
  12. Map Representation
  13. Keyframe Selection
  14. Coarse-to-Fine Visual Localization
  15. Global Localization
  16. ...and 15 more sections

Figures (4)

  • Figure 1: Pipeline of LiteVLoc. The mapping phase builds a topo-metric map from a set of images with poses (Section \ref{['sec:mapping']}). Selected cameras are marked in red; otherwise, they are marked in blue. The map includes two levels, with edges marked in black: level-$1$ for planning and level-$2$ for VLoc. The localization phase estimates camera's poses in a coarse-to-fine manner (Section \ref{['sec:visual_localization']}). The map and real-time localization results can be applied to various applications such as image goal navigation (Section \ref{['sec:navigation']}). The image to illustrate PnP is derived from sheffer2020pnp.
  • Figure 2: Sample images and the distribution of camera poses in the proposed dataset for evaluating map-free relocalization methods.
  • Figure 3: Real-world experiment with a legged robot, with the red curve showning the robot's trajectory estimated by our VLoc method. The nodes (visualized as axes: red-X, green-Y, and blue-Z) and connected edges indicate the topo-metric map's structure. The robot is guided by a sequence of goal images that were captured by the AR glasses (right). It starts inside a room, navigates outside, follows a circular route, and returns. It totally traverses $94.5m$ in $314s$ with an average speed of $0.3m/s$. Green boxes A-E highlight key planning events along the route: (A) a narrow pathway, (B) descending slope, (C and D) outdoor pathways, and (E) the lab region. Refer to the video for a full demonstration of the navigation process.
  • Figure 4: Estimated trajectories on a real-world sequence traversing a bridge. VLoc's trajectory closely aligns with the ground truth, demonstrating high accuracy. The sudden changes in both RobotOdom and RobotOdom-PS are due to corrections from VLoc's results.