GeoLocSFT: Efficient Visual Geolocation via Supervised Fine-Tuning of Multimodal Foundation Models

TL;DR

GeoLocSFT demonstrates that targeted supervised fine-tuning of a large multimodal model with a small, high-quality dataset can achieve state-of-the-art visual geolocation performance, especially in sparsely populated regions.

Contribution

The paper introduces GeoLocSFT, a novel framework that uses supervised fine-tuning of Gemma 3 with limited data to improve geolocation accuracy significantly.

Findings

Supervised fine-tuning with only 2700 image-GPS pairs improves geolocation performance.

GeoLocSFT outperforms baseline models on multiple benchmarks including Im2GPS-3k, YFCC-4k, and MR40k.

Core performance gains are achieved at the fine-tuning stage without complex inference strategies.

Abstract

Accurately determining the geographic location where a single image was taken, visual geolocation, remains a formidable challenge due to the planet's vastness and the deceptive similarity among distant locations. We introduce GeoLocSFT, a framework that demonstrates how targeted supervised fine-tuning (SFT) of a large multimodal foundation model (Gemma 3) using a small, high-quality dataset can yield highly competitive geolocation performance. GeoLocSFT is trained with only 2700 carefully selected image-GPS pairs from our geographically diverse MR600k dataset. Despite this limited data, our SFT-centric approach substantially improves over baseline models and achieves robust results on standard benchmarks such as Im2GPS-3k and YFCC-4k, as well as on our newly proposed and challenging MR40k benchmark, aimed specifically at sparsely populated regions. Further, we explore multi-candidate inference and aggregation strategies but find that the core gains are already realized at the SFT stage. Our findings highlight the power of high-quality supervision and efficient SFT for planet-scale image geolocation, especially when compared to prior methods that require massive databases or complex pipelines. To foster further research, we publicly release the MR40k benchmark dataset.