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

5.2.2RCE2 primary contributions

(Reviewed Thu 24th a.m. plenary GJS & JRO)

JCTVC-O0065 RCE2: C2 - Entropy Coding Compression Efficiency for High Bit Depths [K. Sharman, N. Saunders, J. Gamei (Sony)]

First aspect: The modifications are applied in section 9.3.3.9 Binarization process for coeff_abs_level_remaining, replacing

cRiceParam = Min( cLastRiceParam + ( cLastAbsLevel > ( 3 * ( 1  <<  cLastRiceParam ) ) ? 1 : 0 ), 4 )

by

cRiceParam = Min( cLastRiceParam + ( cLastAbsLevel >> ( 2 + cLastRiceParam ) ), 9 )

Second aspect: In HEVC, the Rice parameter is reset to 0 at the beginning of each 4×4 sub-block. This assumes that the Rice parameter is increasing within the sub-block and then, goes back to the minimum at the beginning of the next sub-block. This assumption might not hold for lossless coding (when transform is bypassed) or for transform skip.

This second aspect sets the value of the Rice parameter at the beginning of the sub-block depending on the Rice value at the end of the previous sub-block in the same TU. If the transform is skipped or bypassed, the value is initialized to the previous Rice minus 1 (not less than 0). Otherwise, the value is initialized to the previous Rice minus 2 (not less than 0).

Decision: Adopt (both aspects, with enable flag in SPS).
Plan CE for D1 (with separate_statistics_flag equal to 0 only) and possibly O0327 / O0129.
Overlapping notes below
Subtest A: Rice parameter update

A1: Gain in normal bit depths lossy (0/0.3%) & lossless (2.5/1.9%) for normal/screen(4:4:4) content, gain for 12/14/16 bit 3/11/23%

A2: Some losses in normal bit depths lossy & lossless, gain for 12/14/16 bit 3/12/25%

Subtest B: Rice parameter initialization

B1: Small gain in normal bit depths lossy & lossless only for screen content, gain for 12/14/16 bit 5/17/28%

B2: Small gain in normal bit depths lossy & lossless only for screen content, gain for 12/14/16 bit 2/3/3%

Subtest C: Rice parameter adaptation

C1: Gain in normal bit depths lossy (0/1%) & lossless (3.3/3.6%) for normal/screen(4:4:4) content, gain for 12/14/16 bit 7/18/28%.

C2: Gain similar to C1

D1: Combination/unification of methods C1 and C2, which performance similar

From the results shown, A1 and D1 are the most relevant results

Update of maximum Rice parameter (as in A1) already gives most of the gain.

C/D are more complex than A, requires operating of 4 counters, separate statistics parameter conveyed at SPS (in case of separate statistics disabled, only 2 counters)

Also A has some variants: With/without reset, and several max values of Rice parameter.

The results shown were without reset and with max=4 for lossy coding, max=7 for lossless and max=9 for high bit depth. It was verbally expressed by the proponents of JCTVC-O0206 that max=9 would also be appropriate for the lossless (normal bit depth) case.

Further discussed Wed 30th (GJS). Decision: To be consistent with other aspects of 4:2:0 12 and 4:2:2 profiles, do not allow the enabling flag to be equal to 1 in those profiles.