VGG-CAE: Unsupervised Visual Place Recognition Using VGG16-Based Convolutional Autoencoder

Abstract

Listen

Translate article

Visual Place Recognition (VPR) is a challenging task in Visual Simultaneous Localization and Mapping (VSLAM), which expects to find out paired images corresponding to the same place in different conditions. Although most methods based on Convolutional Neural Network (CNN) perform well, they require a large number of annotated images for supervised training, which is time and energy consuming. Thus, to train the CNN in an unsupervised way and achieve better performance, we propose a new place recognition method in this paper. We design a VGG16-based Convolutional Autoencoder (VGG-CAE), which uses the features outputted by VGG16 as the label of images. In this case, VGG-CAE learns the latent representation from the label of images and improves the robustness against appearance and viewpoint variation. When deploying VGG-CAE, features are extracted from query images and reference images with post-processing, the Cosine similarities of features are calculated respectively and a matrix for feature matching is formed accordingly. To verify the performance of our method, we conducted experiments with several public datasets, showing our method achieves competitive results comparing to existing approaches.