In order to illustrate the performa

In order to illustrate the performance on London Eye test set, we visualize false positive and false negative image pairs of test data. To achieve that, we extract feature vectors from layer fc7 of the two branches and calculate Euclidean distance between them over test data and sort them in ascending order for negative pairs and in descending order for positive pairs. The visualization is presented in Figure 5. It shows hard negatives and hard positives image pairs encountered by sHybridNet on London Eye test data. Hard positives are examples of positive pairs with largest pairwise feature distances returned on test data by using sHybridCNN feature. Similarly, hard negatives are examples of negative pairs with smallest distances. By looking at the examples in Figure 5, we observe that ground truth labels are imperfect and actually most of the negative pairs with smallest distances represent the same scene and should have been labeled as positive (i.e. matching).