MAPS: Multi-Anchor Projection Similarity for Joint Vision-Language Geo-Localization

Traditional geo-localization models often treat visual and textual queries as independent references, calculating similarity through separate pairwise comparisons. This approach fails to capture the joint semantic subspace created when vision and language are used together to describe a specific location. The authors investigate how to better model this joint query structure to improve retrieval accuracy in vision-language geo-localization (VLGL).

To address the limitations of pairwise alignment, the authors propose the Unified Multi-Anchor Geo-localization (UniMAG) framework. UniMAG embeds visual queries, textual queries, and candidate locations into a shared high-dimensional space. Within this space, the authors introduce Multi-Anchor Projection Similarity (MAPS). Instead of calculating cosine similarity for each modality separately, MAPS constructs an 'anchor plane' from the visual and textual features. It then evaluates candidate locations based on their projection length onto this plane, effectively measuring how well a candidate aligns with the combined semantic constraint. To ensure the model learns features consistent with this geometry, the authors also introduce a MAPS-based contrastive loss that penalizes candidates that deviate from the query-defined anchor plane.

The authors demonstrate that treating VLGL as a multi-anchor geometric alignment problem is superior to standard pairwise methods. By explicitly modeling the relationship between the visual-textual query subspace and the candidate features, the MAPS metric provides a more discriminative ranking criterion. Experimental results on the CORE and CVG-Text datasets show that the UniMAG framework consistently outperforms existing methods, with the most significant gains achieved when MAPS is utilized for both the retrieval ranking and the training objective.