IM2City

Abstract

This study investigated multi-modal learning as a stand-alone solution to image geo-localization problems. Based on the successful trials on the contrastive language-image pre-training (CLIP) model, we developed GEo-localization Multi-modal (GEM) models, which not only learn the visual features from input images, but also integrate the labels with corresponding geo-location context to generate textual features, which in turn are fused with the visual features for image geo-localization. We demonstrated that simply utilizing the image itself and appropriate contextualized prompts (i.e., mechanisms to integrate labels with geo-location context as textural features) is effective for global image geo-localization, which traditionally requires large amounts of geo-tagged images for image matching. Moreover, due to the integration of natural language, our GEM models are able to learn spatial proximity of geo-contextualized labels (i.e., their spatial closeness), which is often neglected by classification-based geo-localization methods. In particular, the proposed Zero-shot GEM model (i.e., geo-contextualized prompt tuning on CLIP) outperforms the state-of-the-art model - Individual Scene Networks (ISN), obtaining 4.1% and 49.5% accuracy improvements on the two benchmark datasets, IM2GPS3k and Place Plus 2.0 (i.e., 22k street view images across 56 cities worldwide), respectively. In addition, our proposed Linear-probing GEM model (i.e., CLIP's image encoder linearly trained on street view images) outperforms ISN even more significantly, obtaining 16.8% and 71.0% accuracy improvements, respectively. By exploring optimal geographic scales (e.g., city-level vs. country-level), training datasets (street view images vs. random online images), and pre-trained models (e.g., ResNet vs. CLIP for linearly probing), this research sheds light on integrating textural features with visual features for image geo-localization and beyond.