Image Retrieval-Based Localization Under Seasonal Changes

In this paper, we achieve the image retrieval-based localization under seasonal changes by learning appearance invariant representation of location with deep neural network. Specifically, we construct the Siamese network architecture to perform the similarity learning between different seasonal images. For each branch of the network, there are three main components. First, a full convolutional network is utilized as local feature extractor to generate rich local features. Then, the NetVLAD layer is utilized to aggregate local features into a global feature. Finally, an fully connected layer is utilized to reduce the feature dimensionality. During the training phase, the weighted soft margin ranking loss is introduced mainly for the convergence acceleration. To show the performance of our method, we have done comparative experiments on the NordLand dataset. The quantitative and qualitative results show that our learning-based method obviously outperforms the traditional hand-crafted methods under seasonal changes.