A Self-Adaptive Feature Extraction Method for Aerial-view Geo-localization.

Area of Science:

• Computer Vision
• Geographic Information Systems
• Machine Learning

Background:

• Cross-view geo-localization matches geographic locations across different image perspectives (e.g., drone vs. satellite).
• Significant scale variations in target buildings due to camera angles, heights, and real-world object sizes pose challenges.
• Existing methods often overlook scale invariance and feature alignment, hindering robust performance.

Purpose of the Study:

• To develop a network capable of mining discriminative representations resistant to scale variations.
• To enhance feature alignment and extract scale-invariant features in a self-adaptive manner for improved geo-localization.

Main Methods:

• Proposed Self-Adaptive Feature Extraction Network (Safe-Net) for end-to-end learning.
• Implemented a global representation-guided feature alignment module using affine transformations.
• Introduced a saliency-guided feature partition module to adaptively process image regions without manual annotations.

Main Results:

• Achieved state-of-the-art performance on University-1652 and SUES-200 benchmarks.
• Demonstrated significant scale adaptive capability, particularly for images with small target buildings.
• Successfully extracted robust feature representations for challenging geo-localization scenarios.

Conclusions:

• Safe-Net effectively addresses scale variations in cross-view geo-localization.
• The proposed modules enable robust feature extraction and alignment.
• The method offers a promising solution for accurate aerial-view geo-localization.