Reading: CNNMC Yokoyama ICCE’20 — Enhanced CNNMC (HEVC Intra Prediction)

Enhanced CNNMC (2018 CVPRW) With Higher Prediction Accuracy

Sik-Ho Tsang
3 min readMay 27, 2020
CNN based Mode Coding (CNNMC) Network Architecture

In this story, “CNN Based Optimal Intra Prediction Mode Estimation in Video Coding”, by Waseda University, and Japan Broadcasting Coporation (NHK), is briefly presented. I just call it CNNMC Yokoyama ICCE’20, since it improves the CNNMC that is published in 2018 CVPRW. I read this because I work on video coding research. Since the authors are different from the one in 2018 CVPRW, I write a new story for it here. (Please read CNNMC first.) This is a paper in 2020 ICCE. (Sik-Ho Tsang @ Medium)

Outline

  1. CNNMC Architecture & MPM Generation Variants
  2. Experimental Results

1. CNNMC Architecture & MPM Generation Variants

1.1. Architecture Variants

CNNMC Architecture Variants (Table I is the table at the top in this story)
  • In this paper, nine CNNMC variants are tried, with different kernel sizes or number of channels in layers conv1 to conv4, as shown above.
  • conv5 is common to all structures.
  • e.g.: Structure (1–2) uses half number of channels as in the table at the top.

1.2. MPM Generation Variants

MPM Generation Variants
  • Nine MPM generation variants are also tried.
  • e.g.: (f): 3 MPMs generated by CNN, and 1 MPM generated by the conventional codec HM.

Based on the above two kinds of variants, accuracy is measured.

2. Experimental Results

Accuracy of CNNMC Architecture Variants
  • As for combination (a), the accuracy is low, it is claimed that it is difficult to estimate the optimal intra prediction modes by solely CNN.
  • As for combination (h), it obtains the highest accuracy.
  • Among (h), structure (2–1) has the accuracy of 0.9188, which is the highest.
  • In the future, this network will be merged into HEVC or even VVC codecs.

During the days of coronavirus, Challenges of writing 30 and 35 stories again for this month have been accomplished. Let me challenge 40 stories!! This is the 39th story in this month.. Thanks for visiting my story..

--

--

Sik-Ho Tsang
Sik-Ho Tsang

Written by Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: https://linktr.ee/shtsang for Twitter, LinkedIn, etc.

No responses yet