Low-complexity merge candidate decision for fast HEVC encoding

1

Low-complexity merge candidate decision for fast HEVC encodingMultimedia and Expo Workshops (ICMEW), 2013 IEEE International Conference on

Muchen LI, Keiichi Chono, Satoshi Goto

2 outline

Introduction Low-complexity merge candidate decision

A. GMV Based Merge Candidate Decision B. Position-Priority Based Merge Candidate Decision

Simulation Results

3 Introduction #1

Although the Merge mode brings significant coding efficiency improvements, the complexity of its Merge candidate decision associated with Rate-Distortion cost computation is proportional to the number of Merge candidates.

The investigation shows that redundant and unnecessary RD cost computation for particular Merge candidates between Ground-truth MVs (GMVs) obtained by motion estimation (ME) and Merge candidates.

4 Introduction #2

For each PU, the Merge mode utilizes up to 5 MVP candidates and selects the best one based on computationally expensive Rate-Distortion (R-D) cost check for each Merge candidate. Sun of absolute transformed difference (SATD) 8-tap filter

5 Introduction #3

Merge mode allows at most 5 Merge MV candidates from the previously inter-coded blocks located in the positions described. 5 spatial 2 temporal

6 Introduction #4

7 Introduction #5

As well as Merge mode, AMVP mode also introduces MVP competition technology and predicts the MV of current PU as one of the MVPs. Ground-truth MV

8 Introduction #6

9 Introduction #7

In the study, it is found that the Merge candidates are typically distributed in the SR and the Merge candidate decision is based on the R-D cost that also uses the SATD measurement.

Checking the SATD for all the Merge candidates is redundant and excessive in some cases.

10 Low-complexity merge candidate decision #A1GMV Based Merge Candidate Decision

The ME in the AMVP mode check and the merge candidate decision employ the same SATD measurement and there is a correlation between them.

Basically, the GMV has the minimum SATD in the SR and it can be taken as a reference for the Merge candidate decision.



In Case (2), the GMV is identical to the BMMV, including the prediction direction, reference picture index and MV components. Case (2) can also be called as equal case.

The GMV represents the true MV of the current PU better than or equal to the BMMV (up to 88.38% for LP and 72.91% for RA).


14 Low-complexity merge candidate decision #B1Position-Priority Based Merge Candidate Decision


A1, B1, and T are always the dominant positions for all the cases. When the CU size becomes larger, the total proportion of A1, B1 and T becomes smaller and the order of them is varying a little for different partitions. A1, B1 and T are available A1, B1 and T are not available


17 Simulation Results #1




Low-complexity merge candidate decision for fast HEVC encoding

Documents