Hierarchical feature alignment for cross-view geo-localization.

Area of Science:

• Computer Vision
• Geographic Information Systems
• Machine Learning

Background:

• Cross-view geo-localization matches street-view images to aerial databases for geographic coordinate estimation.
• Extreme viewpoint and scale variations pose significant challenges to accurate matching.
• Existing methods struggle with stable extraction of global structures and local details.

Purpose of the Study:

• To propose GeoAlignNet (GANet), a novel framework for robust cross-scale and cross-view representation alignment.
• To enhance retrieval-based geo-localization accuracy under challenging conditions.
• To improve the stability and discriminability of feature representations.

Main Methods:

• Developed a hierarchical feature learning framework, GeoAlignNet (GANet).
• Introduced the Spatial Structure Attention (SSA) module for view-invariant spatial layout capture.
• Implemented the Local Representation Refinement (LRP) module for fine-grained local feature modeling.
• Utilized a hybrid objective for joint intra-class compactness and inter-class separability.

Main Results:

• GANet demonstrated competitive performance against existing cross-view geo-localization methods.
• The SSA module effectively captures view-invariant spatial structures and mitigates misalignment.
• The LRP module enhances perception of geometric textures and stability against variations.
• The hybrid objective facilitated stable optimization under extreme structural variations.

Conclusions:

• GANet provides a robust solution for cross-view geo-localization.
• The proposed framework shows strong effectiveness and generalization capabilities.
• GANet advances the state-of-the-art in retrieval-based geo-localization.