Improvements of Image Retrieval-based Visual Localization Using Structured Database

In this paper, we propose a method to improve image retrieve for visual localization by structuring the database. We are studying cloud-based positioning infrastructure system that we call Universal Map. It can reduce various cost as compared with the conventional technique. However, it takes time to estimate the position because the retrieval process is performed from a large amount of images in the database. To solve this problem, we reduce the retrieval time by structuring the database. We designed feature vector representing each image in the database and classified them using clustering method called K-means. We also made virtual sensing image and measured the Euclidean distance to each cluster in order to evaluate the classification results. As a result, the correct cluster was selected up to the third closest cluster. Therefore, we could reduce the retrieval time to 20% so far.