Street View Image Retrieval with Average Pooling Features

Abstract

Street view image retrieval is a challenging subject due to the complexity of the street view. Such images contain a large number of similarly structured buildings and are obscured by other objects such as pedestrians and vehicles. In this paper, we present a street view image retrieval system based on average pooling features and SIFT (Scale-invariant feature transform), which extract location information of query images from retrieval results. The average pooling features extracted from the middle layer of the convolutional neural networks (CNNs) are used to retrieve on a large-scale street view dataset firstly, then the SIFT Re-ranking is performing on the results of former retrieval to improve precision. Besides, before performing retrieval panorama street view images are projected into the perspective view images that are similar to the query image captured by mobile. Finally, 82.14% of images are correctly retrieved through our system and the Re-ranking method increases retrieval performance by about 12.5%. The results show that the method can effectively locate the mobile phone image captured in the city by retrieving the large-scale street view image dataset.