Flow Chart

In 3D-HEVC, the wedgelet searching process in the depth map coding consumes a lot of time. We propose an algorithm in this work to balance the trade-off between coding efficiency and computational complexity using deep learning.

Figure 1: Flowchart for Proposed Algorithm

Figure 2: Detailed Flowchart for Proposed Algorithm

Description

step 1 Get the Luma pixel values from one depth block. (The block can be of size 8x8, 16x16, 32x32)

step 2 Feed the 2D matrix of Luma pixels into learned model for getting the top-16 predictions.

step 3 Add top 16 predictions into the RMD LIST.

step 4 Check whether mode 2 is inside RMD LIST. If yes, add mode 34 into RMD LIST; otherwise jump to step 5.

step 5 Add mode 0, 1, DMM1, DMM4 into RMD LIST.

step 6 Do RMD. For DMM1, only check the directions covered by top-16 predictions.

step 7 Add two modes into FULL RDO LIST. Do full RDO.

step 8 Obtain the best mode for the depth block.

Note

The above process can be applied to a batch of blocks, in which case the time cost of prediction can be optimized. For details, see Time Cost of TF in C++

[Deprecated]Flow chart

This chart has been deprecated. Kept here only for reference.

Deprecation Summary

• For below reason 1 and reason 2, we remove edge strength analysis;
• For below reason 3, we remove the implementation to texture.

Reasons

1. Edge strength analysis is not innovative.
2. Besides, removing it from the flow chart only will decrease the accuracy of ResNet prediction by roughly 2%~3%.
3. And according to Dr.Tsang, since we are only using luma pixel values, it seems we should not apply our model into the texture blocks.

Figure D-1: Flowchart for Proposed Fast Intra Mode Decision Algorithm