Title:Large Scale Landmark Recognition via Deep Metric Learning

Abstract:This paper presents a novel approach for landmark recognition in images that we've successfully deployed at Mail ru. This method enables us to recognize famous places, buildings, monuments, and other landmarks in user photos. The main challenge lies in the fact that it's very complicated to give a precise definition of what is and what is not a landmark. Some buildings, statues and natural objects are landmarks; others are not. There's also no database with a fairly large number of landmarks to train a recognition model. A key feature of using landmark recognition in a production environment is that the number of photos containing landmarks is extremely small. This is why the model should have a very low false positive rate as well as high recognition accuracy.
We propose a metric learning-based approach that successfully deals with existing challenges and efficiently handles a large number of landmarks. Our method uses a deep neural network and requires a single pass inference that makes it fast to use in production. We also describe an algorithm for cleaning landmarks database which is essential for training a metric learning model. We provide an in-depth description of basic components of our method like neural network architecture, the learning strategy, and the features of our metric learning approach. We show the results of proposed solutions in tests that emulate the distribution of photos with and without landmarks from a user collection. We compare our method with others during these tests. The described system has been deployed as a part of a photo recognition solution at Cloud Mail ru, which is the photo sharing and storage service at Mail ru Group.