Back to Search Document details
3rd Meeting: Geneva, May 2016 2016-05-26 07:29
Exploration Experiments on Coding Tools Report
Authors: Elena Alshina Jill Boyce Y.-W. Huang Seung Hwan Kim Li Zhang
Abstract
Seven experiments on coding tools were agreed to carry out between JVET-B and JVET-C meetings in order to get better understanding of technologies considered for inclusion to the next version of JEM, analyse and verify their performance, complexity and interaction with existing JEM tools. This report summarises the status of each experiment.
Meeting Notes
JVET-C0010 Exploration Experiments on Coding Tools Report [E. Alshina, J. Boyce, Y.-W. Huang, S.-H. Kim, L. Zhang]

Summary of Exploration Experiments.

#

Main test and sub-tests

Document

Y-BD-rate (Enc/DecTime)

Cross-check

2.1

Quad-tree plus binary-tree (QTBT) (*)

SW released at April, 19, modified during EE.

JVET-C0024

AI: −3.3% (ET 5.4, DT 1.0)

RA: −3.8% (ET 2.?, DT 1.?)

LD: −4.5% (ET 2.4, DT 1.1)

LDP: −4.5% (ET 2.2, DT 1.2)

JVET-C0056 Samsung

  • Low-complexity Intra configuration

AI: −2% (ET 2.5, DT 1.0)

2.2

Non Square TU Partitioning(**)

SW released and unchanged since April, 19

JVET-C0077

JVET-B0047

AI: −1.5% (ET 1.9, DT 1.0)

RA: −1.0% (ET 1.1, DT 1.0)

LD: −0.7% (ET 1.1, DT 1.0)

LDP: −0.8% (ET 1.1, DT 1.0)

JVET-B0068

Sony

2.3

NSST and PDPC index coding

(all modifications enabled)

JVET-C0042

AI: −0.6% (ET 1.5, DT 1.0)

RA: −0.2% (ET 1.1, DT 1.0)

LD: −0.1% (ET 1.1, DT 1.0)

LDP: −0.1% (ET 1.2, DT 1.0)

JVET-C0059 Samsung

JVET-C0087 Qualcomm

  • w/o removing PDPC restriction

AI: −0.2% (ET 0.9, DT 1.0)

RA: −0.1% (ET 1.0, DT 1.0)

LD: −0.0% (ET 1.1, DT 1.1)

LDP: −0.0% (ET 1.0, DT 1.0)

2.4

De-quantization and scaling for next generation containers

SW released and unchanged since April 19

JVET-C0095

(registered May 25)

2.5

Improvements on adaptive loop filter

SW released and unchanged since April, 19

JVET-C0038

AI: −1.0% (ET 1.0, DT 1.1)

RA: −1.2% (ET 1.0, DT 1.0)

LD: −1.1% (ET 1.0, DT 1.0)

LDP: −1.5% (ET 1.0, DT 1.0)

JVET-C0036

Huawei

JVET-C0057

Samsung

JVET-C0074 Sharp

JVET-C0091

Intel

  • W/o chroma filter vs whole package

AI: 0.1% (ET 1.0, DT 1.0)

RA: 0.0% (ET 1.0, DT 1.0)

LD: 0.0% (ET 1.0, DT 1.0)

LDP: 0.0% (ET 1.0, DT 1.0)

  • W/o prediction from fixed filters vs whole package

AI: 0.3% (ET 1.0, DT 1.0)

RA: 0.2% (ET 1.0, DT 1.0)

LD: 0.1% (ET 1.0, DT 1.0)

LDP: 0.1% (ET 1.0, DT 1.0)

2.6

Modification of Merge candidate derivation

SW released and unchanged since April, 19

JVET-C0035

RA: −0.1% (ET 1.0, DT 1.0)

LD: −0.2% (ET 1.0, DT 1.0)

LDP: −0.2% (ET 1.0, DT 1.0)

JVET-C0060 Samsung

JVET-C0073 Sharp

JVET-C0085

Huawei

  • ATMVP simplification

RA: −0.0% (ET 1.0, DT 1.0)

LD: −0.0% (ET 1.0, DT 1.0)

LDP: −0.0% (ET 1.0, DT 1.0)

  • Merge pruning

RA: −0.1% (ET 1.0, DT 1.0)

LD: −0.2% (ET 1.0, DT 1.0)

LDP: −0.2% (ET 1.0, DT 1.0)

2.7

TU-level non-separable secondary transform (***)

SW released at April, 19, modified during EE.

JVET-C0053

AI: −0.5% (ET 0.8, DT 1.0)

RA: ?% (ET ?, DT ?)

LD: −0.1% (ET 1.0, DT 1.0)

LDP: −0.0% (ET 1.0, DT 1.0)

JVET-C0058 Samsung

JVET-C0086 Sharp

JVET-C0076

Orange, B-com

  • W/o HyGT

AI: −0.1% (ET 0.8, DT 1.0)

RA: ?% (ET 1.0, DT 1.0)

LD: −0.1% (ET 1.0, DT 1.0)

LDP: −0.0% (ET 1.0, DT 1.0)

  • All coeff. sub-groups use secondary transform

AI: −0.5% (ET 0.8, DT 1.0)

RA: ?% (ET ?, DT ?)

LD: −0.1% (ET 1.0, DT 1.0)

LDP: −0.1% (ET 1.0, DT 1.0)

  • Secondary transform is applied for all non-zero TUs (default 2 non zero coeff)

AI: −0.2% (ET 0.8, DT 1.0)

RA: ?% (ET ?, DT ?)

LD: −0.1% (ET 1.0, DT 1.0)

LDP: −0.1% (ET 1.0, DT 1.0)

  • Secondary transform is applied for transform-skip and LM mode

AI: −0.5% (ET 0.8, DT 1.0)

RA: ?% (ET ?, DT ?)

LD: −0.1% (ET 1.0, DT 1.0)

LDP: −0.1% (ET 1.0, DT 1.0)

  • CU-level signalling with HyGT

AI: −0.6% (ET 1.5, DT 1.1)

RA: ?% (ET ?, DT ?)

LD: −0.1% (ET 1.1, DT 1.0)

LDP: −0.0% (ET 1.1, DT 1.0)

Comments:

(*) Full tests data available in cross-check report (not in original contribution), significant chroma gain is observed (~5% AI, LDB and LDP, ~8.5% in RA), significant Class F gain (not included in the previous average number by the CTC template) is observed (AI Y: ~4%, UV: ~7%, RA Y: ~5%, UV: ~9%, LDB Y: ~8%, UV: ~10%, LDP Y: ~8%

(**) Tested vs HM16.6.

(***) Luma BD-rate gain is accompanied by chroma drop. Only partial test data available by May 24.

EE1: QTBT: Gain is slightly higher with other tools off (using QTBT with HM), Has significant increase in encoder runtime, particularly for AI

EE2: Gives some evidence how much of the QTBT gain comes from non-square transform

EE3: NSST/PDPC: Most gain is obtained via removing the PDPC restriction. NSST gives about 0.2%, but is not increasing the complexity

EE4: Dequant: Late document, further review needed

EE5: ALF modifications provide gain without change in encoding/decoding runtime. Modification of chroma filter gives only small benefit.

EE6: No loss by ATMVP simplification; the second aspect avoids duplicate merge candidate, which gives a small gain.

EE7: Secondary transform (hypercubic Givens transform) Results (RA) not fully available yet. AI provides most gain (0.5% on average). While there is gain in luma, some losses occur in chroma in some cases. Reduction of run time because TU level operation does not require a second prediction.

EE1: Quad-tree plus binary-tree (QTBT)

References:
JVET-B0047 JVET-B0051 JVET-B0059 JVET-B1001 JVET-C0024 JVET-C0035 JVET-C0038 JVET-C0042 JVET-C0053 JVET-C0056 JVET-C0058 JVET-C0059 JVET-C0060 JVET-C0066 JVET-C0077
Cited:
JVET-C0001 JVET-D0010 JVET-E0010 JVET-F0010 JVET-K0029 JVET-K1029
ARCHIVE v3 :
JVET-C0010_r2.docx
Decisions
EE1: Quad-tree plus binary-tree (QTBT)
Citation