Optimizing Video Quality Estimation Across Resolutions

A Hitchhiker's Guide to Structural Similarity

“…However, as discussed earlier, using windowed means requires large windows (k ≈ 50, 80) while providing only a minor performance improvement over the baseline. In addition, because CoV pooling performed consistently well across all databases and does not have any hyperparameters to tune, we recommend using spatial CoV pooling on picture or video frame quality maps, and the standard arithmetic (2,1) 0.687 0.705 0.129 FNS 0.692 0.583 0.128 DW (1/8) 0.732 0.715 0.121 Mink (4) 0.755 0.747 0.117 (50) 0.718 0.799 0.880 0.682 0.921 0.882 W-GM (55) 0.719 0.804 0.882 0.684 0.866 0.884 W-HM (78) 0.704 0.809 0.884 0.684 0.842 0.889 DW (1/4) 0.698 0.804 0.888 0.680 0.923 0.887 Mink (1/4) 0 (50) 0.684 0.765 0.863 0.632 0.896 0.868 W-GM (55) 0.683 0.764 0.865 0.633 0.863 0.869 W-HM (78) 0.684 0.765 0.868 0.633 0.839 0.872 DW (8) 0.683 0.768 0.870 0.634 0.910 0.873 Mink (8) 0 As in previous sections, we repeated the experiments on compression distorted data, and reported the results in Tables 14 and 15. Even when we restricted the distortion types, we did not obtain concordant values of hyperparameters across the video databases.…”

“…between the HD source and rendered videos. We propose a suite of algorithms, called Scaled SSIM in [50], which predict SSIM by only using SSIM values computed at the lower compression resolution during runtime. The video compression pipeline in which we solve the Scaled SSIM problem is illustrated in Fig.…”