Deep SIMBAD: Landmark Ranking-based Scene Descriptor for Highly Compressive Self-localization Across Domains

Landmark-based robot self-localization has recently garnered interest as an efficient maintenance-free approach for performing visual place recognition (VPR) across domains (e.g., time of day, weather, and season). In this research, we formulate this landmark-based VPR as a deep learning extension of SIMBAD, and propose a new extremely compact scene descriptor based on landmark ranking. Our contributions are as follows. In our study, a scene is described compactly as an ordered list of IDs of most similar landmarks. Such a descriptor can be efficiently indexed and retrieved by an incremental inverted index. Similarity between query and database scenes can be evaluated discriminatively by using reciprocal rank features (RRFs). The result shows an extremely compact data structure that compresses the descriptors into incremental inverted indexes. Experiments using the public NCLT dataset validated the effectiveness of the proposed approach.