Of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11

JCTVC-M0381 Crosscheck of JCTVC-M0323 proposed by Qualcomm [J. Kim (LG)] [late] [miss]
JCTVC-M0330 Screen Content Coding with Multi-stage Base Color and Index Map Representation [W. Zhu (Beijing University of Technology), J. Xu (Microsoft), W. Ding (Beijing University of Technology)]

(Discussion chaired by D. Flynn.)

This contribution presents a new mode for coding screen content, which is based on the intra coding framework of the high efficiency video coding (HEVC) standard. The scheme first decomposes each intra prediction unit (PU) into color components and structure components. Then a two-stage prediction scheme is employed to generate two prediction indexes for each sample to code those structure components. Experimental results show that on average (including class F and sequences in 4:4:4 format) −34.4%, −24.5% and −31.5% BD-rate saving can be achieved for AI, LD and RA cases.

This is a palette based method for intra coding, wherein a palette derived from an input block is transmitted and a map of pels to palette entries is transmitted known as the structure map. The mode is conditionally enabled when the intra prediction mode is planar, in this case an additional syntax element is used to decide if the mode should be used or if the true planar mode should be used. Compression of the structure map is achieved using a two-stage prediction scheme. Individual entries of the structure map are predicted at the decoder by examining which prediction direction would have been best for the previously reconstructed pel.

	BD-Bitrate Saving(All Intra)			BD-Bitrate Saving(Lowdelay)			BD-Bitrate Saving(RandomAccess)
Sequence	Y	U	V	Y	U	V	Y	U	V
Average (inc L0303)	−34.4%	−30.1%	−30.5%	−24.5%	−22.4%	−22.5%	−31.5%	−27.5%	−28.3%
Average (vs. L0303)	−12.0%	−6.3%	−6.4%	−7.2%	−2.8%	−2.7%	−9.6%	−3.1%	−2.9%

All-Intra Main	(M0330 vs. L0303)			(M0330 + L0303 vs. HM)
	Y	U	V	Y	U	V
Class F	−15.6%	−9.2%	−8.8%	−15.7%	−9.2%	−8.8%
SC (YUV)	−11.7%	−5.8%	−6.0%	−46.6%	−42.8%	−43.6%
SC (RGB)	0.0%	0.0%	−0.1%	0.0%	0.0%	−0.1%
Enc Time[%]	164.6%			176.7%
Dec Time[%]	113.4%			91.2%

The results presented are a combination of this scheme and the dictionary coder of JCTVC-L0303. When compared against the dictionary coder, the gain is −12% BD-rate, the greatest contribution of which comes from sequences where the dictionary coder does not perform well, such as Class F. Other proposals have yielded much lower coding gains when coding the Class F sequences which were captured using analogue methods, however, this proposal seems to retain coding efficiency in this case.

Coding configuration used the range extensions software in AI/RA/LB configurations (adjusted for 4:2:0 as necessary).

Unfortunately the results presented in this contribution are not directly comparable with those of L0323 due to differences in the test configuration.

The contributor recommends further study. It may be worth studying in conjunction with L0323 in some AHG activity. There is no cross-check report.

To be studied in CE with M0350 and M0323 (coordinated by L. Guo).