Iterative finetuning for image geolocalization based on proximity graph segmentation

Image geolocalization using social media photos remains a challenging task due to variations in image content and geographical ambiguity. To improve classification accuracy, we propose an iterative framework that integrates proximity graph-based region segmentation and fine-tuning of Vision Transformer (ViT). First, we construct a proximity graph, where nodes represent photo locations and edges are weighted by the cosine similarity of ViT-extracted image features. The Louvain method is then applied to partition the graph into communities, which serve as class labels for training. After fine-tuning ViT on these labels, the extracted image features are updated, leading to a refinement of edge weights and a re-partitioning of the graph. This iterative process continues, ensuring that the graph partitions increasingly align with visual similarity. Our experimental results demonstrate that this refinement enhances region segmentation and improves geolocalization accuracy compared to conventional methods such as CPlaNet.