Fast Online-Learning Parameters Decision Algorithm
Based on Bayesian Decision Rule for 3D-HEVC
Yayong Li, Xingang Liu, Tao Yu, Yongyong Mei and Peicheng Wang
School of Electronic Engineering
University of Electronic Science and Technology of China
Chengdu, China
liyayong163@163.com
Abstract—3D-HEVC, as the extension of High Efficiency Video
Coding standard, achieves a significant improvement in the coding
efficiency of 3D videos, compared with the Multi-View Video
Coding (MVC). However the improvement causes a great
computational complexity. In this paper, a fast coding parameters
decision algorithm is proposed to reduce the computational
complexity. The process of the selection of Coding Unit (CU),
Prediction Unit (PU) and Transform Unit (TU) are modeled as the
online-learning classification process based on the Bayesian
decision rule. Through the training of the selected feature vectors,
the classifiers can precisely predict the prediction mode of PUs and
whether or not the current CU and TU should be partitioned. The
experimental results show that the proposed algorithm can achieve
about 56% time reduction with a slight RD degradation.
Keywords—3D-HEVC; Bayesian decision rule; fast coding
parameters decision;
I. INTRODUCTION
Due to the immersive viewing experience, three-dimensional
videos are becoming more and more popular, which has been
pushing the advance of the 3D technology. Joint Collaborative
Team on 3D Video Coding Extension Development (JCT-3V)
was established in 2012, and proposed 3D-HEVC standard in
2013. The standard, based on the HEVC standard, is aimed at
achieving a higher compression rate to the multi-view
stereoscopic videos with little degradation in video quality. 3D-
HEVC not only takes full advantages of the spatiotemporal
correlation, but also exploits the inter-view correlation to
remove the redundant, which improves the coding efficiency
more.
Because of the special characteristics of 3D videos, 3D-
HEVC develops some new techniques. For the dependent views,
disparity-compensated prediction (DCP), Inter-View Motion
Parameters Prediction (IVMP), Inter-View Residual Prediction
(IVRP) are introduced to remove the inter-view redundant.
DCP adds the encoded inter-view frames at the same moment
to the reference frame list to reduce the redundant. IVMP
utilizes the inter-view motion information to predict the motion
parameters of the current view. IVRP, similar to the IVMP,
utilizes the inter-view residual to predict the residual of current
view. Since depth map has quite different characteristics than
the texture videos, which consists of large quantity of smooth
regions separated by sharp edges, 3D-HEVC introduces Depth
Modeling Modes(DMM), which can preserve the edge
information better.
For texture videos and depth maps, as showed in Figure 1,
3D-HEVC adopts the same quad-tree coding structure. Coding
Tree Unit (CTU) is still the basic coding unit. One CTU can be
split into four Coding Trees (CU), and one CU can be
recursively split into four sub-CUs in rate distortion
optimization (RDO) process. The maximum size of CU is 64x64
and minimum size is 8x8. Prediction Unit (PU) carries the
prediction information, and all of the prediction modes,
including Skip, 2Nx2N, 2NxN, Nx2N, NxN, 2NxnU, 2NxnD,
nRx2N and nLx2N, intra2Nx2N, intraNxN, PCM must be
checked to find the best mode. It should be noted that NxN mode
is only checked when the current CU is the smallest CU.
Transform Unit (TU) is the basic unit of transform and
quantization. TU, similar to CU, can be split in a quad-tree
structure and each split TU can be further split recursively,
whose maximum size is 32x32 and minimum size is 4x4.
The rest of the papers are organized as follows. Section II
presents the related works about 3D-HEVC. The process of
modeling the parameters selection is showed in section III.
Section IV presents the proposed algorithm, and the experiment
results are showed in the section V. Finally, section VI presents
the conclusion.
II. RELATED WORK
In 3D-HEVC, CU, TU partition process and PU selection
process cost most of the encoding time, and these processes have
a significant effect on the compression rate and the video quality,
so how to predict the best CU, PU and TU in advance is very
important for reducing the encoding time efficiently.
For CU partition process, 3D-HEVC has to recursively split
the current CU into four sub-CU and calculate the RD cost to
find the most optimal CU depth. In order to decrease the
Fig. 1 Example of CU, PU, and TU split structure