Back to Search Document details
30th Meeting: Antalya, TR, April 2023 2023-04-22 12:53
Non-EE2: Bi-predictive IBC for natural and screen content
Authors: yoshitaka kidani KDDI H. Kato KDDI K. Kawamura KDDI
Abstract
This contribution proposes a bi-predictive intra block copy (IBC) to enhance the coding performance of IBC for natural and screen content. EE2-1.8 (JVET-ADxxxx) of this meeting tests IBC for natural content. In the EE2-1.8, VVC spec, and current ECM design, IBC generates prediction samples with only one block vector (BV), i.e., uni-predictive IBC, but it still has room to improve the prediction accuracy of IBC. Therefore, this contribution adds IBC with two BVs, i.e., bi-predictive IBC, besides uni-predictive IBC. The proposed method consists of two types of bi-predictive IBCs. Method 1, naming IBC BVP-merge mode, derives the two BVs from the existing IBC BVP mode and IBC merge mode in ECM-8.0. Method 2, naming bi-predictive IBC merge mode, derives the two BVs from IBC merge candidate list, utilizing two IBC merge indices.
Meeting Notes
JVET-AD0134 Non-EE2: Bi-predictive IBC for natural and screen content [Y. Kidani, H. Kato, K. Kawamura (KDDI)]

This contribution proposes a bi-predictive intra block copy (IBC) to enhance the coding performance of IBC for natural and screen content. EE2-1.8 and EE2-1.9 (JVET-AD0208) of this meeting tests IBC for natural content. In the EE2-1.8/1.9, VVC spec, and current ECM design, IBC generates prediction samples with only one block vector (BV), i.e., uni-predictive IBC, but it still has room to improve the prediction accuracy of IBC. Therefore, this contribution adds IBC with two BVs, i.e., bi-predictive IBC, besides uni-predictive IBC. The proposed method consists of two types of bi-predictive IBCs. Method 1, naming IBC BVP-merge mode, derives the two BVs from the existing IBC BVP mode and IBC merge mode in ECM-8.0. Method 2, naming bi-predictive IBC merge mode, derives the two BVs from IBC merge candidate list, utilizing two IBC merge indices.

The experimental results of the two methods implemented on the top of EE2-1.8a (without fractional IBC) and EE2-1.8c (with fractional IBC) in AI are reportedly {for Y, U, V, EncT, DecT}:

Test 1: IBC BVP-merge mode + bi-predictive IBC merge mode in EE2-1.8a

  • Overall :{-0.51%, -0.59%, -0.62%, 109%, 100%} over EE2-1.8a
  • TGM :{-0.51%, -0.19%, -0.17%, 103%, 98%} over ECM-8.0 / EE2-1.8a

Test 2: IBC BVP-merge mode + bi-predictive IBC merge mode in EE2-1.8c

  • Overall :{%, %, %, %, %} over EE2-1.8c

Test 3: IBC BVP-merge mode over EE2-1.8a

  • Overall :{-0.21%, -0.30%, -0.28%, 101%, 100%} over EE2-1.8a

Test 4: IBC BVP-merge mode over EE2-1.8c

  • Overall :{%, %, %, %, %} over EE2-1.8c

Test 5: Test 2 + encoder optimization

  • Overall :{%, %, %, %, %} over EE2-1.8c

It was agreed to investigate this in an EE.

Encoder runtime significantly increased for method 2, reduction should be investigated. For method 2, it is also recommended to investigate the benefit of bi-prediction for IBC independent of the option of using IBC in both partitions of GPM.

Cited:
JVET-AD0313 X
ARCHIVE :
JVET-AD0134-v2.docx JVET-AD0134-v3.docx JVET-AD0134-v3_clean.docx JVET-AD0134.docx JVET-AD0134.pptx Test1_over_EE2-1.8a.xlsm Test2_over_EE2-1.8c.xlsm Test3_over_EE2-1.8a.xlsm Test4_over_EE2-1.8c.xlsm Test5_over_EE2-1.8c.xlsm
Decisions
Encoder runtime significantly increased for method 2, reduction should be investigated. For method 2, it is also recommended to investigate the benefit of bi-prediction for IBC independent of the option of using IBC in both partitions of GPM.
Citation