Cross-View Image Set Geo-Localization

TL;DR

This work introduces Set-CVGL, a cross-view geo-localization paradigm that uses unordered sets of ground images from diverse viewpoints to locate a query against satellite references. It provides SetVL-480K, a large-scale benchmark with dense ground-to-satellite correspondences, and FlexGeo, a flexible model with two novel components: the Similarity-guided Feature Fuser (SFF) for adaptive, non-sequence-based feature fusion, and the Individual-level Attributes Learner (IAL) for geo-attribute supervised learning. FlexGeo achieves state-of-the-art performance on SetVL-480K and competitive results on SeqGeo and KITTI-CVL, with improvements over 22% in accuracy when using multiple query images. The approach demonstrates the practical value of leveraging multi-view cues for robust geo-localization, aligning with human visual strategies and enabling more reliable localization in varied environments.

Abstract

Cross-view geo-localization (CVGL) has been widely applied in fields such as robotic navigation and augmented reality. Existing approaches primarily use single images or fixed-view image sequences as queries, which limits perspective diversity. In contrast, when humans determine their location visually, they typically move around to gather multiple perspectives. This behavior suggests that integrating diverse visual cues can improve geo-localization reliability. Therefore, we propose a novel task: Cross-View Image Set Geo-Localization (Set-CVGL), which gathers multiple images with diverse perspectives as a query set for localization. To support this task, we introduce SetVL-480K, a benchmark comprising 480,000 ground images captured worldwide and their corresponding satellite images, with each satellite image corresponds to an average of 40 ground images from varied perspectives and locations. Furthermore, we propose FlexGeo, a flexible method designed for Set-CVGL that can also adapt to single-image and image-sequence inputs. FlexGeo includes two key modules: the Similarity-guided Feature Fuser (SFF), which adaptively fuses image features without prior content dependency, and the Individual-level Attributes Learner (IAL), leveraging geo-attributes of each image for comprehensive scene perception. FlexGeo consistently outperforms existing methods on SetVL-480K and two public datasets, SeqGeo and KITTI-CVL, achieving a localization accuracy improvement of over 22% on SetVL-480K.

Figures (8)

• Figure 1: Illustration of the Set-CVGL Task: Using multiple ground images captured from diverse perspectives to match geo-tagged references for location determination.
• Figure 2: Distribution of query data across different datasets. The top part illustrates query image distribution in Sequence-CVGL datasets, while the bottom part shows the distribution in SetVL-480K. The blue line represents the vehicle's route, with blue dots marking sampling points and orange markers indicating the direction of image capture.
• Figure 3: (1) Overview of the proposed FlexGeo method, featuring two modules: IAL and SFF. (2) Illustration of the FlexGeo inference phase. (3) Detailed illustration of the SFF module. (4) Detailed illustration of the IAL module.
• Figure 4: Evaluation of the number of query images. The blue line represents localization results using the SFF module for feature fusion, while the green line shows results with average pooling.
• Figure 5: Heatmap visualization of SetVL-480K dataset images generated by the FlexGeo model. The first column contains satellite imagery and its heatmap. The second and third columns are the heatmaps of four corresponding ground images. The sampling positions and orientations of the ground images are marked on the satellite images.