VI.CONCLUSIONWe have evaluated the performance of Siamese network features for image matching on landmark datasets. There are several conclusions that we can get from our experiments. First, this network architecture is able to learn from such data when one uses pretrained CNN from a related image classification problem as a starting point. We also showed that our approach has promising results of generalization on unseen landmark datasets. We also observed that potentially the imperfect ground truth labels during training are preventing the network to learn and generalize optimally. Nevertheless, it allows to suggest that Siamese architecture together with contrastive loss objective is a good choice for learning features for image matching and retrieval tasks. Moreover, using addi- tional relevant datasets [26], [27] during training might further enhance the accuracy and performance of the approach.