Data for Using Deep Learning to Quantify the Beauty of Outdoor Places

In order to predict the scenic ratings of images for which we do not already have crowdsourced data, we use a transfer learning approach to leverage the knowledge of the Places365 CNN [1], which can predict the place category of a scene with a high degree of accuracy. We modify the Places CNN to instead predict the scenicness of an image. We fine-tune our CNN using 80% of the Scenic-Or-Not dataset [2], and use the remaining 20% test set to check our prediction accuracy. We calculate a performance measure using the Kendall Rank correlation between the predicted scenic scores and the actual scenic scores. The Scenic CNN trained using the Visual Geometry Group (VGG) convolutional neural network architecture delivers the best performance with an overall prediction accuracy of 0.658. We predict the scenicness of images of London uploaded to Geograph (http://www.geograph.org.uk/). This dataset includes all the scenic predictions used to create Figure6 "Predictions of scenic ratings for London images". [1] Zhou, B., Khosla, A., Lapedriza, A., Torralba, A., & Oliva, A. 2016 Places: An image database for deep scene understanding. arXiv preprint arXiv:1610.02055.