Hashing for localization: A review of recent advances and future trends

This paper reviews state-of-the-art hashing-for-localization (HfL) technologies. It begins by reviewing visual localization and hashing-based visual retrieval technologies. In the field of visual localization, camera pose estimation and image geolocalization are two major research topics. Although their application areas differ, they share common key technologies in terms of image feature extraction and matching. Additionally, both face common challenges, including complex feature representation and slow localization speed. Hashing-based visual retrieval is a technology for accelerating retrieval using hashing algorithms. This technology performs feature matching by computing the Hamming distance between compact hash codes, which simplifies computational complexity and improves retrieval efficiency. Therefore, the development of hashing algorithms for accelerating visual localization represents an efficient and robust solution.

This paper introduces a series of HfL technologies that leverage hashing algorithms to accelerate visual localization. The typical process for implementing HfL begins by extracting visual features from a query image for localization and from reference images with localization information. These features are then mapped into compact binary hash codes using trained hash functions while preserving their visual similarities. During localization, the query image, represented by the generated hash code, is compared with the hash codes of the reference images in the database using the Hamming distance. Localization is achieved by selecting the closest matches through ranking, i.e., the localization information of the reference images whose hash codes have small Hamming distances to the query image’s hash code is considered the localization result. This paper reviews four types of HfL technologies: remote sensing HfL, sketch HfL, street view image geolocalization, and encrypting hashing against localization. It also discusses their empirical advantages over existing visual localization methods.

Through methodological and experimental analysis, it is observed that HfL offers the following three main advantages: (a) HfL significantly reduces storage requirements by mapping image features to compact binary hash codes; (b) HfL accelerates the feature matching process by computing the Hamming distance between hash codes; (c) HfL, aided by geographic cluster extraction, achieves exact localization. HfL effectively overcomes the inefficiencies of traditional visual localization technologies and represents a breakthrough in fast localization. In the future, fusing various data sources, including visual, global navigation satellite system, Wi-Fi, and acoustic signals, to enhance HfL will be a key trend. Leveraging large language models to update HfL with semantic information will be another key trend. HfL will provide efficient and accurate localization support across small-, medium-, and large-scale application scenarios.