JVPR: Bilateral Information Interaction by Joint Training for Sequence-Based Visual Place Recognition

Visual Place Recognition (VPR) focuses on retrieving the most similar imagery from database given a query, and it usually takes a video stream as input. Facing omnipresent sequential information in robotics and autonomous systems, neither image descriptor-based sequence match methods nor sequential descriptor aggregation techniques consider correlation between image descriptors and sequential descriptors. Instead, they are processed independently as two separate information flows. In this work, we propose a novel joint-training visual place recognition (JVPR) to enhance the consistency and robustness between two kinds of descriptors, thereby coupling former two independent pipelines and maintaining the robustness of both image and sequential descriptors. We also pre-process recently released Mapillary Street-Level Sequences dataset so that more choices in sequence-based VPR researches are afterwards available to the community, and we benchmark our JVPR as well as SeqNet on several representative cities.