Towards location recognition using range images

Retrieving the location of a mobile device by matching a query image to a database of geo-tagged imagery is one popular application of content-based image retrieval (CBIR). Standard CBIR-based approaches exploit appearance features of the environment for the matching process. Many locations, however, are characterized by distinct structural (geometric) features. We investigate whether a standard appearance-based CBIR pipeline can be adapted to perform location retrieval using a range image-based representation of the environment. The contributions are three-fold: We design a rigorous experimental setup using an extensive and challenging indoor dataset. Secondly, we compare the state-of-the-art feature algorithm specifically designed for range images, the Normal Aligned Radial Feature (NARF) [1], against some of the most established appearance-based features. Thirdly, we combine the high key point detection rate of NARF, with the robustness of the Speeded-Up Robust Feature for range-image based location recognition. This detector-descriptor combination, which we coin NURF, leads to 15% improvement in absolute location recognition performance compared to simple NARF in our experimental setup.