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

## TL;DR

This paper improves camera relocalization by enhancing a neural network and pose solver, achieving better accuracy and speed on multiple datasets.

## Contribution

A refined ACE network with skip/dense connections and channel attention, along with modified loss and pose solver for improved performance.

## Key findings

- Average localization errors were reduced by up to 30% compared to the original ACE network.
- Computational times were cut by approximately 10% on tested datasets.
- The approach showed robustness and practicality across 7scenes, 12scenes, and wayspots datasets.

## Abstract

This paper presents an improved approach for scene-aware camera relocalization using RGB images and poses. Building upon the ACE network, we proposed a refined head structure that integrates skip and dense connections alongside channel attention mechanisms. Additionally, we introduced modifications to the loss function and pose solver, leveraging SQPnP and iterative optimization. These enhancements led to significant improvements in the localization accuracy and speed, as evidenced by our experiments on the 7scenes, 12scenes, and wayspots datasets. Here, we show that the average localization errors were reduced by up to 30% and the computational times were cut by approximately 10% compared to the original ACE network, demonstrating the practicality and robustness of our approach.