The ability to have unmanned ground vehicles navigate unmapped off-road terrain has high impact potential in application areas ranging from supply and logistics, to search and rescue, to planetary exploration. To achieve this, robots must be able to estimate the traversability of the terrain they are facing, in order to be able to plan a safe path through rugged terrain. In the work described here, we pursue the idea of fine-tuning a generic visual recognition network to our task and to new environments, but without requiring any manually labelled data. Instead, we present an autonomous data collection method that allows the robot to derive ground truth labels by attempting to traverse a scene and using localization to decide if the traversal was successful. We then present and experimentally evaluate two deep learning architectures that can be used to adapt a pre-trained network to a new environment. We prove that the networks successfully adapt to their new task and environment from a relatively small dataset.