LocalSPED: A Classification Pipeline that Can Learn Local Features for Place Recognition Using a Small Training Set

Abstract

Visual place recognition is a key component for visual-SLAM. The current state-of-art methods use CNNs (Convolutional Neural Networks) to extract either a holistic descriptor or local features from the images. In recent work, a holistic descriptor method with the name SPED was proposed. In this paper, SPED is extended to a local feature configuration called LocalSPED by applying several modifications and by introducing a novel feature pooling method. Several variations of SPED and LocalSPED are trained on a smaller dataset and their performances are evaluated on several benchmark datasets. In the experiments, LocalSPED handles the decreased training set size significantly better than the original SPED approach and provides better place recognition results.

This work is funded by the German Federal Ministry for Economic Affairs and Energy (Bundesministerium für Wirtschaft und Energie) in the project ViPRICE.