Landmark Recognition in Photos

Deep learning has become the most popular approach for solving image retrieval and recognition problems. In particular, significant results were achieved for the task of landmark recognition. However, most existing systems do not solve it directly but use the results of solving the retrieval problem. These solutions achieve high recognition accuracy, but they have longer inference time and use more memory. This work describes a way to build a system for landmark classification, which allows to avoid intermediate stages and is faster during inference. A described system can perform classification without location information - only by photo, however, the presence of location information increases the accuracy of recognition. To train the deep neural network, a dataset containing photos depicting landmarks that Odnoklassniki users uploaded to the network was built. The process of building the dataset is fully automated and does not require additional human assessment. Keywords: deep learning, image classification, landmark recognition.