Figure 7

Dissimilarities between interpolated features and ground truth. 10 networks with \(\gamma =1.0\) have been trained. Dissimilarity is low in case of a no or a single missing feature, and increases with number of missing features up to 40% for six missing features. In general, binary semantic features are inferred with better accuracy than non-binary semantic features. The variance of the different networks for the features ‘height’ and ‘weight’ are highest.