Transfer Learning for Classification of Parking Spots using Residual Networks

The paper proposes a classifier with a residual convolutional architecture for visual parking spot classification into classes "empty" and "occupied". The classifier is trained on the well-known PKLot dataset. Transfer of the resulting model to data with new challenging modalities (such as snow, partially obscured vision, reflections, mist, ...) is tested - to this end a new dataset has been collected by the authors. It is shown that the original classifier fails in some of these unfamiliar settings, but that the failure modes can successfully be corrected using transfer learning. (C) 2019 The Authors. Published by Elsevier B.V.