On-Device Scalable Image-Based Localization via Prioritized Cascade Search and Fast One-Many RANSAC

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Accurate localization is crucial for mobile applications, but current methods often rely on GPS or network connectivity, limiting their use in urban environments.
• Resource constraints on mobile devices pose challenges for computationally intensive localization tasks.
• Existing 2D-3D correspondence methods struggle with repetitive structures common in urban settings.

Purpose of the Study:

• To design an efficient, on-device system for large-scale urban localization using only images.
• To develop a GPS-agnostic and network-independent localization solution for mobile devices.
• To improve the accuracy and speed of 2D-3D correspondence search and pose estimation in complex urban environments.

Main Methods:

• Integration of compact image retrieval and 2D-3D correspondence search for location estimation.
• A novel hashing-based cascade search algorithm for rapid 2D-3D correspondence computation.
• A new one-many RANSAC (Random Sample Consensus) algorithm to handle repetitive structures and enhance pose estimation accuracy.

Main Results:

• The proposed 2D-3D correspondence search method achieves state-of-the-art localization accuracy on benchmark datasets.
• The system demonstrates potential for large-scale urban localization entirely on typical mobile devices, validated on a large Google Street View dataset.
• The one-many RANSAC effectively addresses challenges posed by repetitive building features in localization.

Conclusions:

• The developed on-device system offers a viable solution for GPS-agnostic, large-scale urban localization.
• The novel hashing-based search and one-many RANSAC significantly advance the efficiency and accuracy of image-based localization.
• This research paves the way for ubiquitous and accurate mobile localization in diverse urban environments.