Reproduce “From Coarse to Fine: Robust Hierarchical Localization at Large Scale” on OpenResearch

About this paper

Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, Marcin Dymczyk

A robust hierarchical visual localization system for large-scale environments, developed by researchers at ETH Zürich and Sevensense Robotics AG, integrates global and local feature extraction into a single neural network. This approach achieves real-time 6-DoF pose estimation at over 20 FPS, outperforming efficient baselines by a factor of 10 while maintaining state-of-the-art accuracy on challenging datasets with severe appearance changes.

Key insights

• Learned local and global features demonstrate superior performance compared to hand-crafted features, especially under challenging environmental conditions.
• The hierarchical coarse-to-fine localization paradigm effectively combines the robustness of global descriptors with the precision of local matching, particularly beneficial for large-scale, perceptually diverse environments.
• A single, monolithic network (HF-Net) can efficiently extract both global and local features by maximizing computation sharing, achieving comparable performance to dedicated, larger teacher models.

Summary by alphaXiv.