Reading: Jin PCM’17 — Fast QTBT Partition Algorithm for JVET Intra Coding Based on CNN (Fast VVC Prediction)

“RD Maintaining” Setting: 43.69% Complexity Reduction With Only 0.77% BD-Rate Increase. “Low Complexity” Setting: 62.96% Complexity Reduction With 2.06% BD-Rate Increase.

In this story, Fast QTBT Partition Algorithm for JVET Intra Coding Based on CNN (Jin PCM’17), by Shanghai University, is briefly presented. It is the enhanced version of Jin ICIP’17 that I have presented in this morning. And I’ve just found this paper right now. Thus, I decided to write a story mainly describe the main difference from Jin ICIP’17. (Please feel free to read Jin ICIP’17 first for the details of QTBT in VVC, and also the proposed CNN approach.)

There are two settings provided in this paper where Jin ICIP’17 has not provided, i.e. RD Maintaining and Low Complexity, to obtain different time reduction and coding birate.

This is a paper in 2017 PCM. (Sik-Ho Tsang @ Medium)

Outline

Network Architecture
VVC Implementation
Experimental Results

1. Network Architecture

The network is similar to the Jin ICIP’17 one.
The only difference is that the input QP is also added onto the 64-length feature vector to become 65-length feature vector after global pooling.
But in Jin ICIP’17, the input QP is only added onto the 48-length feature vector.

2. VVC Implementation

The implementation is more or less the same as shown above where 32×32 CU is used as basis. But the figure is much beautiful and clear here.

But there are two settings provided in this paper where Jin ICIP’17 has not provided, i.e. RD Maintaining and Low Complexity.
With the class label predicted, different splitting modes are tried for different settings, as shown above.
Low Complexity setting tests fewer splitting mode to increase the time reduction, but also sacrifice the video quality.

3. Experimental Results

**BD-Rate (BR) (%), Accuracy (AC) (%), Average Time Saving (ATS), and CNN Net Time Consumption (NetT) (%) Compared to JEM-3.1**

As shown in the table above, an averaged complexity reduction of 43.69% under “RD Maintaining” setting with 0.77% negligible BD-rate increase.
Under “Low Complexity” setting, we achieve 62.96% average complexity reduction with 2.06% acceptable BD-rate increase.
“AC” column indicates CU depth decision accuracy within the candidate depth range. 88.46% and 77.08% on average are achieved under “RD Maintaining” and “Low Complexity” setting respectively.
It is important to realize that, when class “3” was misclassified as “4”, the encoder still can fine the optimal partition depth through RDO in the end, since the optimal partition depth still inside the candidate depth range of class “4”.
For “NetT” column, the computational complexity added to the encoder associated with the CNN classifier is account 2.95% and 4.51% on average under “RD Maintaining” and “Low Complexity”.
The time saving of our proposed algorithm is 3.8–5.5 times as much as that of D0077.

During the days of coronavirus, A challenge of writing 30 stories again for this month has been accomplished. A new target of 35 stories is set by now. This is the 31st story in this month.. Thanks for visiting my story..