Active Map-Matching: Teacher-to-Student Knowledge Transfer from Visual-Place-Recognition Model to Next-Best-View Planner for Active Cross-domain Self-localization

Abstract

Training a next-best-view (NBV) planner for active cross-domain self-localization is an important and challenging problem. Unlike typical in-domain settings, the planner can no longer assume a constant environment state, but must regard it as a high-dimensional component of the state variable. This study is motivated by the ability of recent visual place recognition (VPR) techniques in recognizing such a high-dimensional environment state in the presence of domain shifts. Our aim is to transfer the state recognition ability from VPR to NBV. However, such a VPR-to-NBV knowledge transfer is a non-trivial task for which no known solution exists. Herein, we propose the use of a reciprocal rank feature, derived from the field of transfer learning, as dark knowledge for transfer. Specifically, our approach is based on the following two observations: (1) The environment state can be compactly represented by a local map descriptor, which is compatible with typical input formats (e.g., image, point cloud, and graph) of VPR systems, and (2) an arbitrary VPR system (e.g., Bayes filter, image retrieval, and deep neural network) can be modeled as a ranking function. Experiments involving nearest neighbor Q-learning show that our approach can yield a practical NBV planner even under severe domain shifts.