A survey on 3D-LUT performance in 10-bit and 12-bit HDR BT.2100 PQ

In the cinema industry, during production and post-production, the use of 3D-LUTs has been proven to be a flexible and fast solution to perform color transformations, or to give a particular look to a captured scene. Unfortunately, the dimension of the 3D-LUT is limited by the available space in processing hardware, so there must be a tradeoff between the precision of the transformation (to avoid visual artefacts such as banding), and the space required to store it. Nowadays professional equipment, but also commercial televisions, are able to display gamuts wider than DCI-P3 with higher dynamic range. The new standard for HDR television is now ITU-BT.2100 [11], which has a color gamut that is almost double the size of Rec.709. Our aim was to determine the required minimum size of a 3D-LUT in the ITU-BT.2100 color conversion which avoids unnecessary visual artefacts. We defined a metric, mean3std, to use as the JND threshold allowing us to determine when an image is free of visible artifacts due to 3D-LUT interpolation error. — Tetrahedral interpolation outperformed all other interpolations for both SDR and HDR applications. It can achieve the same quality of the Trilinear interpolation using a 3D-LUT 20% to 25% smaller on average. The widely used Trilinear interpolation generated the largest errors among the tested images. In 10-bit SDR, in order to achieve unnoticeable interpolation errors using the trilinear interpolation method, a 3D-LUT larger than 41×41×41 is necessary. However, using tetrahedral interpolation a 31×31×31 3D-LUT is sufficient. In 12-bit HDR, in order to achieve unnoticeable interpolation errors, a 3D-LUT larger than 72×72×72, is required while a 3D-LUT size of 55×55×55 is sufficient if using the tetrahedral interpolation. The saturated and dark images resulted in a much larger minimum size 3D-LUT to achieve a JND<1 pointing to the possibility that ICTCP is likely not suited for dark images. Further studies are needed with emerging new color difference metrics to evaluate the 3D-LUT size for these types of images. Finally, we determined that reducing the bit-depth from 12-bit to 10-bit will require a 3D-LUT 50% larger on average to maintain a comparable image quality.

Published

2018-10

Content type

Original Research

Keywords

HDR, 3D-LUT, ICTCP, DeltaICTCP, ΔICTCP

DOI

10.5594/M001821

ISBN

978-1-61482-960-7