Multi-query Diagnosis of Deep Self-localization Network via Reciprocal Rank Fusion

A series of recent methods of visual robot self-localization are based on training a deep convolutional neural network (CNN) as a visual place classifier (VPC). In this study, we focus on the issue of domain adaptation of such a VPC-CNN in a cross-domain scenario and investigate a new fault-diagnosis-based strategy known as domain-shift localization, where a VPC-user robot aims to localize domain-shifted regions in the robot workspace during its daily navigation. Our approach exploits the fact that the result of a diagnostic query for a VPC-CNN can be generally described by class-specific rank values, and that results from multiple queries can be integrated via reciprocal rank fusion. The effectiveness of the proposed method is verified experimentally using the public NCLT dataset.