An Efficient and Scalable Collection of Fly-inspired Voting Units for Visual Place Recognition in Changing Environments

TL;DR

This paper introduces DrosoNet, a highly compact fly-inspired neural network, combined with a voting mechanism of multiple classifiers, to achieve robust visual place recognition in changing environments with minimal computational resources.

Contribution

The paper presents DrosoNet, a novel extremely compact neural network inspired by fruit flies, and a voting mechanism that enhances robustness and efficiency in visual place recognition tasks.

Findings

DrosoNet achieves state-of-the-art robustness with minimal computational cost.

Voting mechanism improves consistency and accuracy over single classifiers.

Models outperform existing methods on benchmark datasets in accuracy and efficiency.

Abstract

State-of-the-art visual place recognition performance is currently being achieved utilizing deep learning based approaches. Despite the recent efforts in designing lightweight convolutional neural network based models, these can still be too expensive for the most hardware restricted robot applications. Low-overhead VPR techniques would not only enable platforms equipped with low-end, cheap hardware but also reduce computation on more powerful systems, allowing these resources to be allocated for other navigation tasks. In this work, our goal is to provide an algorithm of extreme compactness and efficiency while achieving state-of-the-art robustness to appearance changes and small point-of-view variations. Our first contribution is DrosoNet, an exceptionally compact model inspired by the odor processing abilities of the fruit fly, Drosophyla melanogaster. Our second and main contribution is a voting mechanism that leverages multiple small and efficient classifiers to achieve more robust and consistent VPR compared to a single one. We use DrosoNet as the baseline classifier for the voting mechanism and evaluate our models on five benchmark datasets, assessing moderate to extreme appearance changes and small to moderate viewpoint variations. We then compare the proposed algorithms to state-of-the-art methods, both in terms of precision-recall AUC results and computational efficiency.