Parameter-efficient few-shot geo-localization via low-rank adaptive fine-tuning for cross-view image matching

Current cross-view image matching and localization algorithms have achieved high accuracy, yet no effective solution has been proposed for few-shot datasets in cross-view image matching and localization tasks. To address cross-view image matching and localization under few-shot conditions, this paper proposes a method that fine-tunes large-scale pre-trained models using Low-Rank Adaptation (LoRA) technology. The AM-RADIO model is employed as the visual feature extraction network, and a LoRA-based fine-tuning strategy is adopted to achieve parameter-efficient adaptation. By introducing a low-rank decomposition mechanism, we significantly reduce the number of trainable parameters while preserving the representation capabilities learned from large-scale datasets. This approach substantially decreases computational resource consumption and training time costs. Experimental results demonstrate that the proposed method achieves 51.19% R@1 accuracy and 76.12% R@5 accuracy on the VIGOR few-shot dataset, outperforming the Sample4Geo algorithm by 23.12% and 23.66%, respectively. These results indicate that the AM-RADIO model with LoRA fine-tuning exhibits enhanced feature representation and matching performance in few-shot scenarios. This study provides a new solution for cross-view image matching and localization tasks, particularly suitable for applications where data acquisition is challenging. By integrating large-scale pre-trained models with parameter-efficient fine-tuning strategies, the proposed method successfully realizes image localization under few-shot conditions, advancing the application of visual retrieval technologies in resource-constrained environments.