SeqNet: Learning Descriptors for Sequence-Based Hierarchical Place Recognition

Visual Place Recognition (VPR) is the task of matching current visual imagery from a camera to images stored in reference map environment. While initial VPR systems used simple direct image methods or hand-crafted features, recent work has focused on learning more powerful features and further improving performance through either some form sequential matcher / filter hierarchical process. In both cases single-image based system still far perfect, putting significant pressure sequence (in case systems) pose refinement stages. this paper we present novel hybrid that creates high match hypothesis generator using short learnt descriptors, which enable selective control score aggregation single descriptors. Sequential descriptors are generated temporal convolutional network dubbed SeqNet, encoding sequences 1-D convolutions, then matched against corresponding dataset provide an ordered list place hypotheses. We perform shortlisted separate pipeline produce overall hypothesis. Comprehensive experiments challenging benchmark datasets demonstrate proposed method outperforming state-of-the-art same amount information. Source code supplementary material can be found online. 1 1 [Online]. Available: https://github.com/oravus/seqNet.