Features extracted from the propose

Features extracted from the proposed sHybridCNN have better performance in 4 cases out of 5 compared to original AlexNet and HybridNet respectively. It also confirms the results illustrated in Figure 3 and proves that sHybridCNN can efficiently distinguish positive and negative pairs. That is, sHybridCNN outperforms Hybrid- CNN on images of unseen landmarks.In one case (London Eye) all three methods have almost similar ROC curves and the PR curve of sHybridNet shows lower precision among easy positive pairs (i.e. PR curve drops in the beginning). However, as explained in detail below, deeper analysis indicates that in the London Eye test set there seems to be particularly many image pairs with incorrect ground truth labels.