Indoor-outdoor classification with human accuracies: Image or edge gist?

We investigate the utility of human performance data on indoor-outdoor scene categorization in improving the generalization performance of a machine indoor-outdoor classifier. On 50 indoor and 50 outdoor scenes, the human categorization accuracies are obtained for these stimuli rendered as either real images or line drawings. We study two types of features, image gist and edge gist, which are the scene gist features extracted from the original image and the edge map of the image respectively. Using human accuracies on real images and line drawings as constraints on these two sets of features in training a max-margin classifier, we observe 4% improvement in classifying never seen 10000 indoor and 10000 outdoor images. Our experiments also reveal that edge gist characterizes indoor scenes far better than image gist. Therefore, not only human labeling is necessary for machine classification, but how humans err on the labeling is instrumental for learning better generalizing features and machine classifiers.