Back to Search Document details
6th Meeting: Hobart, March 2017 2017-04-02 09:02
AHG8: On cross-format S-PSNR-NN
Authors: Yong He InterDigital Xiaoyu Xiu InterDigital Y. Ye InterDigital
Abstract
In Geneva, cross-format spherical quality metrics were included into the 360 common test conditions as defined in JVET-E1030. For cross-format quality metrics, the original video and the reconstructed video to be evaluated have different projection formats and/or different resolutions. S-PSNR-NN is one of the cross-format metrics. In order to avoid the impact due to interpolation filters, S-PSNR-NN measures the distortion between samples at nearest neighbor integer positions in the projection plane, instead of interpolating at fractional sample positions. However, when S-PSNR-NN is measured across two projection planes with different formats and/or resolutions, the nearest neighbor rounding operation in each projection plane may cause the reference sample and the reconstructed sample to no longer correspond to the same position on the unit sphere. In such case, cross-format S-PSNR-NN may not provide reliable quality measurement and coding performance comparison. This document proposes a modified S-PSNR-NN calculation method to solve this problem. It is asserted that simulation results validate that the BD rates reported by the modified cross-format S-PSNR-NN are consistent with other metrics, such as end to end S-PSNR-NN and WS-PSNR.
Meeting Notes
JVET-F0042 AHG8: On cross-format S-PSNR-NN [Y. He, X. Xiu, Y. Ye (InterDigital)]

Cross-format spherical quality metrics were included into the 360° common test conditions as defined in JVET-E1030. For cross-format quality metrics, the original video and the reconstructed video to be evaluated have different projection formats and/or different resolutions. S-PSNR-NN is one of the cross-format metrics. In order to avoid the impact due to interpolation filters, S-PSNR-NN measures the distortion between samples at nearest neighbour integer positions in the projection plane, instead of interpolating at fractional sample positions. However, when S-PSNR-NN is measured across two projection planes with different formats and/or resolutions, the nearest neighbour rounding operation in each projection plane may cause the reference sample and the reconstructed sample to no longer correspond to the same position on the unit sphere. In such case, cross-format S-PSNR-NN may not provide a reliable quality measurement and coding performance comparison. This document proposes a modified S-PSNR-NN calculation method to try to solve this problem. It was asserted that simulation results validate that the BD bit rates reported by the modified cross-format S-PSNR-NN are consistent with other metrics, such as end to end S-PSNR-NN and WS-PSNR.

The method is modifying the set of points on the sphere that are compared, such that each maps to an integer sample position in the original. The number of points is around 655K. The set of points compared is the same for each projection format, and interpolation is performed to compare each point against the corresponding fractional sample position in the reconstruction.

It was pointed out that the method is not symmetric, as the results would be different when the original and reconstruction are interchanged (unless they are in the same projection format).

Decision (SW360): Replace the previous cross format S-PSNR-NN by the new method.

JVET-F0042 AHG8: On cross-format S-PSNR-NN [Y. He, X. Xiu, Y. Ye (InterDigital)]

Replace the previous cross format S-PSNR-NN by the new method.

References:
JVET-E1030
Cited:
JVET-F0090 X JVET-F1003 JVET-G1003 JVET-H1004 JVET-K1004 JVET-L1004 JVET-M1004 JVET-Q2004 JVET-S2004 JVET-T2004
ARCHIVE v2 :
JVET-F0042.docx JVET-F0042_presentation.pdf
Decisions
Decision (SW360): Replace the previous cross format S-PSNR-NN by the new method.
Replace the previous cross format S-PSNR-NN by the new method.
Citation