Deep Learning Features at Scale for Visual Place Recognition

TL;DR

This paper trains CNNs specifically for visual place recognition at scale, using a new large dataset, and demonstrates improved performance and insights into learned features compared to generic pre-trained networks.

Contribution

It introduces the SPED dataset for large-scale training and develops a multi-scale feature encoding method for invariant place recognition.

Findings

Achieved 10% performance improvement over existing methods.

Developed a new large-scale dataset (SPED) for training.

Provided insights into network features learned for place recognition.

Abstract

The success of deep learning techniques in the computer vision domain has triggered a range of initial investigations into their utility for visual place recognition, all using generic features from networks that were trained for other types of recognition tasks. In this paper, we train, at large scale, two CNN architectures for the specific place recognition task and employ a multi-scale feature encoding method to generate condition- and viewpoint-invariant features. To enable this training to occur, we have developed a massive Specific PlacEs Dataset (SPED) with hundreds of examples of place appearance change at thousands of different places, as opposed to the semantic place type datasets currently available. This new dataset enables us to set up a training regime that interprets place recognition as a classification problem. We comprehensively evaluate our trained networks on several challenging benchmark place recognition datasets and demonstrate that they achieve an average 10% increase in performance over other place recognition algorithms and pre-trained CNNs. By analyzing the network responses and their differences from pre-trained networks, we provide insights into what a network learns when training for place recognition, and what these results signify for future research in this area.