Re-ranking methods for visual geo-localization with domain shift

Visual Geo-localization is the task of estimating the geographical coordinates where a given photo has been taken. The problem is well known in computer vision literature and the common approach relies on image retrieval technique. Recent works achieved high performances leveraging deep convolutional neural network to embed an image in a fixed low dimensional sized vector. However, we observe that domain shift is still a big challenge and the accuracy of these methods can drop when facing such a challenge, for example, a dataset of query images taken at night. In this work, we explore how re-ranking methods based on spatial verification and deep learning can handle this problem by providing new benchmark with state-of-the-art models on datasets with night queries. Moreover, we introduced a new labeled dataset that contains night query images taken in San Francisco.