# From Coarse to Fine: Robust Hierarchical Localization at Large Scale

**Authors:** Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, Marcin Dymczyk  
**arXiv:** [1812.03506](https://arxiv.org/abs/1812.03506)  
**Date:** 2018-12-09  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.03506v2/

## Key points

- Authors address the need for robust and accurate visual localization in applications like autonomous driving, mobile robotics, and augmented reality
- Proposed HF-Net is a hierarchical localization approach using a monolithic CNN for predicting local features and global descriptors for precise 6-DoF localization
- Key innovation of HF-Net is its coarse-to-fine localization paradigm, reducing runtime requirements and enabling real-time operation
- Demonstrated remarkable robustness in localization across large appearance variations by harnessing learned descriptors
- Sets new state-of-the-art performance on challenging benchmarks for large-scale localization tasks
- Offers a promising solution to visual localization in dynamic environments, enhancing capabilities in autonomous systems and augmented reality applications

## Summary

In their paper titled "From Coarse to Fine: Robust Hierarchical Localization at Large Scale," authors Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk address the crucial need for robust and accurate visual localization in various applications such as autonomous driving, mobile robotics, and augmented reality. They propose HF-Net, a hierarchical localization approach that leverages a monolithic Convolutional Neural Network (CNN) to simultaneously predict local features and global descriptors for precise 6-degree-of-freedom (6-DoF) localization. The key innovation of HF-Net lies in its coarse-to-fine localization paradigm. This strategy reduces runtime requirements and enables real-time operation of the system. By harnessing learned descriptors within their method, the authors demonstrate remarkable robustness in localization across large variations in appearance. Their approach sets a new state-of-the-art performance on two challenging benchmarks for large-scale localization tasks. The proposed HF-Net offers a promising solution to the complex problem of visual localization in dynamic environments and paves the way for enhanced capabilities in autonomous systems and augmented reality applications.