Recognizing visual places from landscapes with zero-shot learning

Image processing and deep learning methods are being developed day by day and the need of recognizing objects and extracting information from visual-based digital multimedia data like pictures and videos is increasing. Therefore, the algorithms are also changing rapidly to meet the requirements of the tasks in everyday life. Zero-shot learning is a new topic and it encourages having small datasets. It still accomplishes the recognition task efficiently. As the name implies, Zero-shot learning is the process of making predictions for the unseen or untrained categories of data based on some amount of data that is learned by the training process earlier. In this study, visual recognition which is based on detecting the city, country, and continent is investigated. There is no or enough similar work on this problem up to now. The unseen places are recognized by training on similar places by categorizing the cities, countries, and continents. Labels and some auxiliary features like a 3D color histogram and GIST features are used to increase the detection accuracy. A new dataset is created for this study and with this work, the importance of the amount of data in the dataset is discussed and various metrics like performance and duration are demonstrated. Also, the comparison with the regular training problem setup is discussed. After all the results are examined, it is seen that ZSL performs better when search space is constrained with only unseen classes at test time. Also, the original generalized ZSL method performs better than the other generalized ZSL method and also inductive ZSL. Compared to traditional learning, various ZSL methods can give sufficient results too.