EfiLoc: large-scale visual indoor localization with efficient correlation between sparse features and 3D points

Important location information of a query image can be obtained directly through indoor 3D points. However, the 3D model-based indoor positioning is still an open issue to be addressed, especially in large-scale dynamic environments. We design and realize the positioning system for large indoor scenes called the EfiLoc. First, we develop a lightweight network model, which can quickly extract discriminative global deep features to improve the discrimination of similar scenes. Another property is that the generated sparser main global descriptors can greatly reduce the retrieval time of multi-dimensional features. Second, we innovatively implement the efficient association of 3D point with the 2D features generated by its projection regions. Preserving the associations of the pixels in some key areas of the image, the precise and quick large-scale indoor localization can be realized. The experimental results show that EfiLoc can achieve good positioning accuracy and is of better robustness to the environment of weak textures and similar scenes compared with current state-of-the-art vision-based solutions.