Organisation internationale de normalisation

5Multi-view coding (MVC)

5.1Core experiment #1 & related docs: Fine motion matching for motion skip mode in MVC

5.1.1.1.1JVT-Z021 ( Prop 2.2) [H. Yang, Y. Chang, J. Huo (Xidian University), S. Lin, S. Gao, L. Xiong (Huawei)] CE1: Fine motion matching for motion skip mode in MVC

At the Shenzhen meeting, a fine motion matching technique was first proposed in JVT-Y037. Some issues about the technique were raised that needed clarification, and a core experiment was set up. This contribution reportedly continues with JVT-Y037, fulfills the mandates of the core experiment plan, and provides results for the proposed technique. The experimental results reportedly show that, compared with the performance of the current MVC JD and JMVM, BD gains of 0.169 dB / 4.11% and 0.083 dB / 2.11% can be obtained as an average over all views; with BD gains of 0.379 dB / 9.19% and 0.181 dB / 4.64% respectively, on average of the views that employs the proposed technique.
Clarification of reported percentage improvement relationships:

• 
  2.11% average gain over all views (for 8 sequences) relative to current motion skip design of JMVM.
  
• 
  4.64% average gain over the views that can use motion skip (for 8 sequences) relative to current motion skip design of JMVM.
  
• 
  4.11% average gain over all views (for 8 sequences) relative to current JD, which does not include motion skip feature.
  
• 
  9.19% average gain over the views that can use motion skip (for 8 sequences) relative to current JD, which does not include motion skip feature.
  

Current design of motion skip in MVC uses a 16-sample GDV (global disparity vector) resolution. Entire GOP uses same GDV value. Contribution proposes to change this to an 8-sample resolution, with a block offset sent to determine the final 16x16 region in the reference picture to use for motion inference. Also send a flag to indicate whether to use a list 0 or list 1 reference picture in this process.

Supports single loop decoding.
Decoder complexity increase: Seems negligible – some extra syntax and minor calculations on MB level.

Encoder complexity increase: Performs, e.g., 9*9*2 = 162 SSE calculations which were not otherwise needed (fast search is also obviously possible).

No change relative to what was proposed in JVT-Y037 (although some simplification of encoding process).
Proposal does not include CAVLC support.
Remark: That seems like a problem (perhaps for motion skip and intensity compensation in general, if CAVLC operation hasn't been tested for them).
Proposal does not include draft text. Considering that the book-keeping that is needed to infer the motion from the new non-macroblock-aligned positioning, the amount of necessary draft text might be substantial.
Remark: Gain seems small.
Remark: Might be good for support of single-loop decoding.
Remark: These results are for multi-loop decoding. How much do you get for single-loop decoding? After some discussion, rough estimates were 19% for multi-loop without new coding tools and 11% for single-loop.
Question: What would be the performance with 8-sample GDV but not adding the refinement? Response: Approximately no improvement from doing that alone – it is the refinement that provides the gain.
Remark: It seems clear that coding tools will not be in MVC phase 1.
Discussed again after further consideration of that issue.
Remark: Phase 1 seems sufficient serve the need of enabling the relevant applications that are ready to be deployed today. With or without low-level decoding tools, these tools still end up with a bit rate that is roughly proportional to the number of views. The market may not be ready yet for further profile definitions without a more major difference than this, e.g., 20% better compression than phase 1 MVC.
Later in the meeting, CAVLC operation was developed by the proponent, with roughly similar relative gain reported. Draft text was also drafted by the proponent.
JVT decision: Adopted into JMVM.
It was further agreed to set up an AHG to investigate MVC enhancements. But it was also agreed that we are in no rush to create more new MVC profiles.
The JMVM software is (already) structured in a way that makes it easy to remove experimental stuff from it. We need to make sure that this stays true.

5.1.1.1.2JVT-Z033 ( Info) [Y. Chen (TUT), Y.-K. Wang (Nokia), S. Liu (USTC), M. M. Hannuksela, H. Li (Nokia)] CE1: Information on motion skip and CE 1

In this contribution, information is given on the performance of the current motion skip in JMVM and the performance and the functionalities of potential tools related to the CE 1. In multiple-loop decoding, the original motion skip reportedly contributes an average bit-saving of around 2%. The method proposed by Huawei in JVT-Z021 reportedly doubles the bit-savings to around 4%.

5.1.1.1.3JVT-Z037-V ( Info) [Y. Su, A. Segall (Sharp)] Verif of JVT-Z021: Fine motion matching for motion skip mode in MVC (CE1)

This document provides a verification report of Huawei's response to CE1 on motion skip mode in MVC. Huawei provided Sharp the software for JVT-Z037. Sharp reportedly carefully inspected the software, compiled the software, and generated results as specified in the CE. All results reportedly matched. JVT-Z021 was thus reported to have been verified.