Semantic Clustering of Image Retrieval Databases used for Visual Localization

In the paper Semantic Clustering of Image Retrieval Databases used for Visual Localization we explore how to partition databases for visual localization into smaller databases using semantic information.

Abstract

Image-retrieval-based visual localization can become computationally expensive on small unmanned aerial systems as reference databases grow. This work organizes database images into smaller clusters using semantic land-cover information, assigns query images to relevant clusters, and restricts retrieval to semantically similar subsets. Experiments with a dedicated aerial semantic-segmentation dataset show that this divide-and-conquer strategy reduces the search space and retrieval time, making large-scale outdoor visual localization more practical for resource-constrained aerial platforms.

Citation

@inproceedings{holzemann2025semantic,
  title        = {Semantic Clustering of Image Retrieval Databases used for Visual Localization},
  author       = {H{\"o}lzemann, Henry and Fiolka, Torsten},
  booktitle    = {2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages        = {6998--7007},
  year         = {2025},
  organization = {IEEE}
}