One of the fundamental problems in footwear forensics is that the distribution of class characteristics in the local population is not currently knowable. Surveillance devices for gathering this data are just half of the battle — it is also necessary to process the data gathered using these devices and identify relevant features. This presentation will describe progress made in automatic identification of relevant footwear features – brand, shoe size, and tread pattern elements, as well as complications which arise when combining machine learning algorithms with human-friendly features. Using transfer learning to connect pre-trained neural networks to newly gathered and labeled training data, this method bridges the gap between unfriendly numerical features and descriptors used by examiners in practice. Leveraging both clean training data and “messy” data gathered from the local community using newly developed footwear surveillance devices, the authors will present developments in footwear forensics which will enable examiners to testify as to the frequency of class characteristics in the local population in the very near future.