CNN-based method with self-supervision for visual place recognition

An open problem in the artificial intelligence community is built an algorithm to able geo-localize a given photo, overcoming the multiple problems related to the domain shift between the images used during the training and the ones passed at test time. During this thesis, our contribution didn’t just focus on research but we have also implemented a software that is easy to use for any kind of user, as well as creating a dataset that can be used for further research. In order to exploit the VPR problem with a deep learning method, we have used the current state-of-the-art CNN called NetVLAD [1], that we have properly modified to speed up the process and for studying the results on our dataset. In particular, we have tested the final network with a third-domain dataset and tried also a self-supervision approach, to make the network more confident with the different domains belong to training and testing phases. Moreover, we have also partially investigated the effect of some artificial occlusions on the image that needs to be geo-localize, seeing if they can be useful to focus the network attention on the relevant object inside a photo, instead of the dynamic ones. Finally, two kind of software are developed, the first one is composed by a set of steps, to download and create all the necessary stuff related to the dataset that a user want to create, while the second one is a graphical user interface, used to upload a photo and visualize the network results.