B.Dataset and data preprocessingAs described in section III-B, in our experiments we use a set of images of 5 different landmarks (London Eye, San Marco, Tate Modern, Times Square and Trafalgar) downloaded from Flickr. To analyse generalization performance of the proposed algorithm, we should evaluate sHybridCNN on an unseen data which is not presented in training data at all. Therefore, we construct a training dataset as a combination of image pairs of 4 landmarks and a test dataset as a set of image pairs of the remaining landmark. Following this procedure, we get 5 different test and training datasets. The images and the list of positive image pairs for each landmark were originally provided by [2]. In addition, we randomly generated negative pairs utilizing images of the same landmark, so the number of matched and unmatched pairs in test datasets is equal. In contrast to test data, all training datasets are unbalanced. Specifically, the number of dissimilar pairs in training data is 1.5larger than the number of similar image pairs.
B.Dataset and data preprocessingAs described in section III-B, in our experiments we use a set of images of 5 different landmarks (London Eye, San Marco, Tate Modern, Times Square and Trafalgar) downloaded from Flickr. To analyse generalization performance of the proposed algorithm, we should evaluate sHybridCNN on an unseen data which is not presented in training data at all. Therefore, we construct a training dataset as a combination of image pairs of 4 landmarks and a test dataset as a set of image pairs of the remaining landmark. Following this procedure, we get 5 different test and training datasets. The images and the list of positive image pairs for each landmark were originally provided by [2]. In addition, we randomly generated negative pairs utilizing images of the same landmark, so the number of matched and unmatched pairs in test datasets is equal. In contrast to test data, all training datasets are unbalanced. Specifically, the number of dissimilar pairs in training data is 1.5larger than the number of similar image pairs.<br>
正在翻译中..