An Empirical Study of Global Descriptors for Image-based Localization in Dense Urban Scenes

Abstract

Background: Visual place recognition is an interesting technology that can be used in many domains such as localizing historical photos, autonomous navigation and augmented reality. The main stream of research in that domain was based on the use of local invariant features like SIFT. Little attention was given to region descriptors which can encompass local and global visual appearances. In this paper, we provide an empirical study on two main visual descriptors: (i) Local Binary Patterns, and (ii) Covariance matrices. Methods: In order to enhance the discriminative power of the final descriptor of each type, multi-block based descriptors are designed and compared. The descriptor corresponding to each input is formed by the concatenation of the features extracted from each building block. We show experimental results on matching test images with reference images acquired in dense urban scenes in the streets of the city of Paris. The problems of scale changes and occlusions are both treated by simulation. Different combinations of processing steps are treated (covariance and LBP descriptors, mono and multi-blocks, L1 and Chi-Squared distances). Results: The obtained results for the tested scenarios show an improvement in the classification rate for at least three scenarios when using multi-block based features rather than mono-block based ones. Conclusion: The use of multi-block based features can thus enhance the discrimination of the obtained final descriptor. The corresponding matching algorithms can lead to both high accuracy and scalability. Keywords: Global image descriptors, image-based localization, image retrieval, image matching.