1 of 14

LDL: Line Distance Functions for Panoramic Localization

Junho Kim, Changwoon Choi, Hojun Jang, and Young Min Kim

Dept. of ECE, Seoul National University

2 of 14

Motivation

  • LDL localizes a panorama image with respect to a point cloud

3D Point Cloud

Query Image

6DoF Localization

Applications

AR Glasses

Robots

3 of 14

Motivation

  • LDL localizes a panorama image with respect to a point cloud

3D Point Cloud

Query Image

6DoF Localization

Why panoramas?

  • Wide FoV: Lesser ambiguities
  • Hardware becoming cheaper & accessible

vs

Panorama

Regular Fov

Ricoh Theta

Meta Aria Glasses

2x 150° cameras

4 of 14

Motivation

  • LDL localizes a panorama image with respect to a point cloud
  • LDL exploits lines in 2D and 3D to enhance robustness and reduce map size during localization

3D Line Cloud

2D Line Segments

3D Point Cloud

Query Image

6DoF Localization

  • Smaller map size
  • Robust to 2D-3D domain gaps
  • Robust to Illumination changes

5 of 14

Overview of LDL

  • LDL performs localization through a three-step process

Lines, Principal Directions,

and Local Features

1. Input Preparation

3. Pose Refinement

Local Feature Matching

2. Coarse Pose Search

3D Line Distance Functions

2D Line Distance Functions

Matching

6 of 14

Overview of LDL

  • Lines and principal directions are extracted for coarse pose search
  • Local feature descriptors are extracted for pose refinement

Lines, Principal Directions,

and Local Features

1. Input Preparation

2. Coarse Pose Search

3D Line Distance Functions

2D Line Distance Functions

Matching

3. Pose Refinement

Local Feature Matching

Input Panorama

Line Segments

Local Features

Principal Directions

7 of 14

Overview of LDL

  • Lines and principal directions are extracted for coarse pose search
  • Local feature descriptors are extracted for pose refinement

Lines, Principal Directions,

and Local Features

1. Input Preparation

2. Coarse Pose Search

3D Line Distance Functions

2D Line Distance Functions

Matching

3. Pose Refinement

Local Feature Matching

Input Point Cloud

Local Features

Line Segments

Principal Directions

8 of 14

Overview of LDL

  • Coarse pose search is performed by comparing line distance functions
  • Line distance functions are defined over the unit sphere for 2D, 3D lines

Lines, Principal Directions,

and Local Features

1. Input Preparation

2. Coarse Pose Search

3D Line Distance Functions

2D Line Distance Functions

Matching

3. Pose Refinement

Local Feature Matching

Line Segments

2D Line Distance Functions

3D Line Map (Top-Down)

3D Line Projections

9 of 14

Overview of LDL

  • Line distance functions are compared for pool of poses within the map
  • Selected poses with similar distance functions are used for refinement

2. Coarse Pose Search

3D Line Distance Functions

2D Line Distance Functions

Matching

2D Line Distance Functions

3D Line Map (Top-Down)

3D Line Distance Functions

Matching

10 of 14

Distance Function Decomposition

  • Line distance functions are decomposed along the principal directions to enhance coarse pose search

2D Line Distance Function Decomposition

3D Line Distance Function Decomposition

Decomposed Distance Function Comparison

Q

11 of 14

Localization Performance (Stanford 2D-3D-S)

Comparison Against Existing Baselines

Legend

PC

PICCOLO (ICCV 2021)

SB

Structure-based method

CD

Chamfer distance-based method

LT

Line Transformer (RA-L 2021)

Runtime per Pose

12 of 14

Conclusion

  • We propose LDL, a line-based panoramic localization algorithm
  • LDL exploits line distance functions for fast and accurate pose search
  • Experiments verify that our method can indeed perform effective line-based localization with a fast runtime

13 of 14

What Will Be the Next Steps?

  • Currently LDL does not incorporate “learning” of any form
  • LDL often fails in large scenes such as hallways
    • Can we leverage neural networks to describe line distributions?
  • Devising neural fields that can effectively describe line maps will be interesting

Difficult Hallway Scenes

Neural Fields for Better Description?

Nearby Line

Information

Spatial Descriptors

14 of 14

Thanks for Listening!