A Guide to Structureless Visual Localization

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Visual localization estimates camera pose in known scenes, crucial for autonomous systems.
• Structure-based methods use 3D models but lack flexibility for scene changes.
• Structureless methods use image databases, offering easier updates but less research focus.

Purpose of the Study:

• To provide the first comprehensive discussion and comparison of structureless visual localization methods.
• To evaluate the performance of structureless approaches against established and recent techniques.

Main Methods:

• Comparative analysis of various structureless visual localization algorithms.
• Extensive experiments evaluating pose accuracy based on geometric reasoning.
• Benchmarking against structure-based and pose regression methods.

Main Results:

• Structureless methods utilizing classical geometric reasoning achieve higher pose accuracy.
• Classical absolute or semi-generalized relative pose estimation outperforms pose regression.
• Structureless methods offer flexibility at a slight cost to pose accuracy compared to structure-based methods.

Conclusions:

• Structureless visual localization is a viable and flexible alternative, especially when scene dynamics are frequent.
• Classical geometric approaches are key to maximizing accuracy in structureless localization.
• Future work should focus on bridging the accuracy gap between structureless and structure-based methods.