LoRetta: New AI model unlocks global-scale satellite image matching

Researchers from Peking University and collaborators have developed LoRetta, a foundation model designed to tackle one of computer vision's toughest challenges: global-scale remote sensing dense image matching. The team addressed the core problem of matching satellite images taken under vastly different conditions (time, season, viewpoint) by reformulating dense matching as a two-step process: first localizing matchable regions and affine geometry, then refining residuals within aligned frames. This approach fundamentally improves reliability when dealing with large geometric offsets and partial overlaps in remote sensing data.

The breakthrough comes with the introduction of LEVIR-GM, a massive benchmark dataset featuring 103,000 aligned image pairs and 827,000 augmented pairs spanning six continents, five years of data, and resolutions from 0.5m to 1024m. On this benchmark, LoRetta achieves an 83.3% area under the curve (AUC) - outperforming the previous state-of-the-art RoMa v2 by 1.6 points while delivering 6.5 and 8.2 point PCK gains at 1 and 2 pixels respectively. Crucially, it reduces inference latency by 47.8%, making it practical for real-time applications. The model's transferability has been validated through astronaut-to-satellite and UAV-to-satellite geolocalization experiments, demonstrating its potential as a reusable geometric aligner across diverse platforms.