Implicit Neural Representation in Visual Place Recognition

Visual Place Recognition is the task of identifying a previously visited location from a new image using visual cues alone. In past years, the main approach was image retrieval, which consists of comparing an image's global descriptor—computed by a deep neural network—and finding the most similar images in a database. It is clear that this method has problems of robustness and dimensionality. Indeed, it requires the query images to be not too different from the database images, and as the region to be learned grows, it demands increasing memory and computation time for the search. In recent years, some methods have been developed to overcome these limitations. This thesis proposes a new method based on the concept of Implicit Neural Regression. The idea is that, since the global descriptor produced by the deep neural network can distinguish images well, we can attempt to create an implicit map in the latent space in which similar images are close and dissimilar ones are far apart, and infer the position of a new, unseen image solely from its position in the latent space. If this is possible, we could obtain a very fast and lightweight method: the position of a query image could be obtained in a single forward pass, and the required memory would be limited to the network’s weights.