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

arXiv:2403.19584v1 Announce Type: new
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 …

abstract arxiv cells classification computer computer vision cs.ai cs.cv database earth foundation grid image images information locations retrieval retrieval-augmented surface type vision