Memorable Maps: A Framework for Re-Defining Places in Visual Place Recognition

Mubariz Zaffar,Klaus D Mcdonald-Maier,Shoaib Ehsan,Michael Milford

doi:10.1109/tits.2020.3001228

Abstract

This paper presents a cognition-inspired agnostic framework for building a map for Visual Place Recognition. This framework draws inspiration from human-memorability, utilizes the traditional image entropy concept and computes the static content in an image; thereby presenting a tri-folded criteria to assess the ‘memorability’ of an image for visual place recognition. A dataset namely ‘ESSEX3IN1’ is created, composed of highly confusing images from indoor, outdoor and natural scenes for analysis. When used in conjunction with state-of-the-art visual place recognition methods, the proposed framework provides significant performance boost to these techniques, as evidenced by results on ESSEX3IN1 and other public datasets.

Highlights

Visual Place Recognition (VPR) is a well-defined but highly challenging module of a Visual-SLAM (Simultaneous Localization and Mapping) based autonomous system [1]
We first show that images collected from the first stage of ESSEX3IN1 lead to poor performance of VPR systems and are not good for insertion into a topological-map
We present the area-under-the Precision-Recall curve (AUC) improvement of different VPR systems when plugged with our framework on all datasets discussed in sub-section IV-A. This is followed-up with a sub-section dedicated to qualitative analysis showing sample images selected and discarded from all datasets

Summary

Introduction

Visual Place Recognition (VPR) is a well-defined but highly challenging module of a Visual-SLAM (Simultaneous Localization and Mapping) based autonomous system [1]. Recent advances in SLAM research can be broken down into semantic mapping (surveyed in [3]) and Visual Place Recognition (surveyed in [1]), where the latter can be annexed into the former [3] This specific work concerns metric, topological and topometric maps having single/multiple images as nodes (landmarks) of the map. For visual place recognition, ‘Places’ have been selected/sampled based on time-interval [4], distance [5] or distinctiveness [6] in different approaches. These approaches are discussed in depth . Due to limited number of images being stored in the map, it is critical to select those images that can be matched successfully upon repeated traversal-the motivation for this research

Methods

Results

Conclusion