Visual Geo-Localization Based on Pretrained Models With Global and Local Features

Visual geo-localization is a critical task of the geolocation field, aiming to match query images with a large database of images from diverse locations. Recent research has shown that pretrained models for general visual learning tasks perform exceptionally well on large datasets, offering effective feature representations for a wide range of visual perception problems. Nevertheless, current models demonstrate constrained generalizability in feature extraction, with their performance in visual grounding tasks deteriorating under complex environmental conditions. This limitation principally arises from three interrelated challenges: 1) sensitivity to illumination fluctuations; 2) vulnerability to perspective alterations; and 3) susceptibility to interference from nondiscriminative visual clutter. To address these issues, we propose a new feature aggregation model that combines features into global descriptors focused on discriminative regions by integrating global relationships within each feature map through a cascade of feature blending. In addition, we introduce attention mechanisms that aggregate global information and emphasize task-relevant regions to guide the rediscrimination of local features. The image similarity score is then derived and reordered based on the optimal number of matching points between the candidate images and the query image. Experimental results demonstrate that our method performs well on large-scale datasets across multiple benchmarks. Furthermore, our approach outperforms single-stage methods using general pretrained models and the benchmark of two-stage methods in most cases.