STEAM: Stable Self-Training with Elastic Matching and Adaptive Purification
STEAM is an unsupervised framework for drone and satellite geo-localization that uses self-training to find GPS-free locations by matching images. It integrates three parts: a stable feature extractor, elastic matching to find good matches, and adaptive purification to keep the matches reliable. This method achieves state-of-the-art accuracy on benchmarks like University-1652 and SUES-200.