Large-scale Visual Geo-localization: Coarse-to-fine Cross-view Image Matching and 3D Spatial Reference Construction

Abstract

Global Navigation Satellite Systems (GNSS) are widely used for precise positioning across many industries. However, their effectiveness is often compromised in GNSS-denied or degraded environments, particularly in urban canyons where dense, tall buildings block satellite signals. In contrast, image-based visual geo-localization functions independently of GNSS, making it a crucial complement for the robust operation of autonomous systems when satellite signals are unavailable. This task typically involves matching images captured from vastly different viewpoints, such as drone versus satellite imagery. Due to the significant disparity in viewpoints, these images exhibit substantial differences in appearance and resolution, making the process highly challenging. 

This talk will introduce a coarse-to-fine pipeline designed to address these difficulties. Specifically, it will discuss how to achieve scene retrieval and correctness reasoning by leveraging multimodal large language models, followed by depth-informed image matching for fine-grained geo-localization. Finally, the presentation will detail the construction of large-scale 3D spatial references to further enhance geo-localization accuracy in complex environments.

Speaker

Dr Zhi Zheng completed his MPhil-PhD program at the School of Remote Sensing and Information Engineering, Wuhan University, in 2023. He is currently a Postdoctoral Fellow at the Chinese University of Hong Kong (CUHK), funded by CUHK’s Research Fellowship Scheme. His research focuses on integrating cutting-edge Artificial Intelligence with satellite remote sensing for cross-view geo-localization, 3D city modelling, and sustainable urban development. He has co-authored approximately 30 papers in peer-reviewed journals and leading conferences, including ISPRS Journal of Photogrammetry and Remote Sensing, International Journal of Applied Earth Observation and Geoinformation, and AAAI. One of his papers was selected as an ESI Highly Cited Paper, and another was featured as a journal frontispiece, both with him as the first author.