Unified and Real-Time Image Geo-Localization via Fine-Grained Overlap Estimation

Area of Science:

• Computer Vision
• Geospatial Artificial Intelligence
• Machine Learning

Background:

• Image geo-localization is challenging due to cross-view and cross-modal differences.
• Real-time, accurate localization in a unified framework remains an open problem.
• Existing methods struggle with the vast differences in modalities and views between source and target platforms.

Purpose of the Study:

• To propose a novel fine-grained overlap estimation method for accurate image geo-localization.
• To develop a unified framework for real-time cross-view and cross-modal image localization.
• To enhance matching accuracy by focusing on salient and subtle overlapping regions.

Main Methods:

• Deep convolutional neural networks extract high-level semantic features.
• A novel overlap scanning module (OSM) mines spatial and channel dependencies for fine-grained overlap identification.
• Triplet ranking loss optimizes the network for precise region matching.

Main Results:

• FOENet demonstrates superior performance on cross-view and cross-modal benchmarks.
• Recall accuracy at top 1 (R@1) shows significant improvement, up to 70.6%.
• The model achieves real-time inference speeds (up to 82.3 FPS) on a single GPU.

Conclusions:

• The proposed FOENet effectively addresses cross-view and cross-modal image geo-localization challenges.
• Fine-grained overlap estimation is crucial for accurate image matching in diverse scenarios.
• FOENet offers a promising solution for real-time, high-accuracy geo-localization systems.