Window-to-window BEV representation learning for limited FoV cross-view geo-localization.

Area of Science:

• Computer Vision
• Geospatial Artificial Intelligence

Background:

• Cross-view geo-localization faces challenges with significant perspective changes and limited ground-view field of view (FoV).
• Unknown ground image orientation and missing camera parameters create ambiguity in matching ground features to Bird's Eye View (BEV) representations.

Purpose of the Study:

• To develop a novel framework for learning Bird's Eye View (BEV) representations directly from ground features, addressing challenges of unknown orientation and limited FoV.
• To improve the accuracy and robustness of cross-view geo-localization systems.

Main Methods:

• Proposed a Window-to-Window BEV (W2W-BEV) representation learning framework.
• Implemented adaptive window-scale matching between BEV queries and ground references.
• Utilized depth-guided BEV initialization to project ground features into BEV space for enhanced matching.

Main Results:

• The W2W-BEV framework adaptively matches BEV queries to ground references at a window scale.
• BEV features are generated by attending to semantically relevant ground features within matched windows.
• Demonstrated significant superiority over state-of-the-art methods on benchmark datasets, particularly under challenging conditions.

Conclusions:

• W2W-BEV effectively bridges the cross-view domain gap by learning from limited FoV ground imagery.
• The proposed method enables learning aligned BEV features by borrowing information from ground views.
• Achieved superior performance in cross-view geo-localization with unknown orientation and limited FoV.