Img2Loc: Revisiting Image Geolocalization using Multi-modality Foundation Models and Image-based Retrieval-Augmented Generation

TL;DR

This work tackles the problem of precise image geolocalization and critiques traditional classification and retrieval approaches. It introduces Img2Loc, a training-free system that leverages CLIP-based image embeddings to build a large image-location database and uses retrieval-augmented prompts with multi-modality foundation models (e.g., GPT-4V, LLaVA) to generate geographic coordinates. By incorporating both similar and dissimilar reference points, Img2Loc achieves superior geolocalization accuracy on Im2GPS3k and YFCC4k without any model fine-tuning. The approach highlights the promise of combining retrieval with generation in foundation models for grounding outputs in external data, offering scalable and robust localization performance. Overall, Img2Loc advances the field by demonstrating a practical, model-lean path to high-precision geolocalization using contemporary multimodal AI.

Abstract

Geolocating precise locations from images presents a challenging problem in computer vision and information retrieval.Traditional methods typically employ either classification, which dividing the Earth surface into grid cells and classifying images accordingly, or retrieval, which identifying locations by matching images with a database of image-location pairs. However, classification-based approaches are limited by the cell size and cannot yield precise predictions, while retrieval-based systems usually suffer from poor search quality and inadequate coverage of the global landscape at varied scale and aggregation levels. To overcome these drawbacks, we present Img2Loc, a novel system that redefines image geolocalization as a text generation task. This is achieved using cutting-edge large multi-modality models like GPT4V or LLaVA with retrieval augmented generation. Img2Loc first employs CLIP-based representations to generate an image-based coordinate query database. It then uniquely combines query results with images itself, forming elaborate prompts customized for LMMs. When tested on benchmark datasets such as Im2GPS3k and YFCC4k, Img2Loc not only surpasses the performance of previous state-of-the-art models but does so without any model training.

Figures (3)

• Figure 1: The image geolocalization problem refers to predicting the coordinates of any given image.
• Figure 2: The architecture of the proposed framework.
• Figure 3: The retrieval and neighbor search pipeline.