RankIQA: Learning from Rankings for No-reference Image Quality Assessment Supplementary Material

Here we provide insight into the ability of our RankIQA approach to learn to discriminate distortions, results illustrating the convergence properties of our fast Siamese backpropagation technique, and experimental results on additional IQA datasets. Learning from Rankings and IQA discrimination. We trained our network on the Places2 [12] dataset until convergence, but performed no fine-tuning on IQA scores. We then plot as histograms the output of our Siamese network on images from the Waterloo [4] dataset distorted with four different distortions as shown in Fig. 1. In the plot, we divide the observations according to the true distortion level (indicated by the color of the histogram). The model discriminates different levels of distortions on Waterloo, even though the acquisition process and the scenes of the two datasets are totally different. Efficient Siamese backpropagation. We compare our method to both standard random pair sampling, and a hardnegative mining method similar to [10]1. The comparison of convergence rate on JP2K, JPEG, GB and GN is shown in Fig. 2. For all four distortions, the efficient Siamese backpropagation not only converges much faster, but also converges to a considerably lower loss depending on how difficult is the specific distortion. It is notable that for the easiest distortion GN, our method converges fast and obtains slightly better performance than hard-negative mining method in the end, however for other three relatively difficult distortions, our method achieves a significant improve-