Enhanced Camera Relocalization Through Optimized Accelerated Coordinate Encoding Network and Pose Solver

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Camera relocalization is crucial for autonomous systems.
• Existing methods like the ACE network have limitations in accuracy and speed.
• Scene-aware approaches require robust feature extraction and pose estimation.

Purpose of the Study:

• To improve scene-aware camera relocalization using RGB images and poses.
• To enhance the ACE network with a refined architecture and optimized solvers.
• To achieve higher localization accuracy and computational efficiency.

Main Methods:

• Proposed a refined network head with skip/dense connections and channel attention.
• Modified the loss function and implemented SQPnP with iterative optimization for pose solving.
• Evaluated the approach on benchmark datasets: 7scenes, 12scenes, and wayspots.

Main Results:

• Achieved up to 30% reduction in average localization errors.
• Reduced computational times by approximately 10% compared to the original ACE network.
• Demonstrated significant improvements in localization accuracy and speed.

Conclusions:

• The proposed enhancements offer a more practical and robust camera relocalization solution.
• The refined approach effectively balances accuracy and computational efficiency.
• This work contributes to advancing scene-aware visual localization techniques.