Back to Search Document details
16th Meeting: Geneva, October 2019 2019-09-27 10:56
CE3-related: MIP downsampling process with re-trained MIP matrix
Abstract
CE3-1.1 (JVET-P0064 [3]) proposed a method of Matrix based Intra Prediction (MIP) downsampling by averaging two samples from the original boundary. The reported simulation results by applying the CE3-1.1 MIP downsampling method and the VTM6.0 MIP matrix showed coding efficiency impact of AI (0.02%) RA (0.01%). We believe that the BDR penalty is due to an inconsistent downsampling process between matrix training and VTM simulation.
JVET-P0401 CE3-related: MIP downsampling process with re-trained MIP matrix [Z. Zhang, K. Andersson, D. Saffar, R. Sjöberg, J. Ström, R. Yu (Ericsson)]

CE3-1.1 (JVET-P0064) proposed a method of Matrix based Intra Prediction (MIP) downsampling by averaging two samples from the original boundary. The reported simulation results by applying the CE3-1.1 MIP downsampling method and the VTM6.0 MIP matrix showed coding efficiency impact of AI (0.02%) RA (0.01%). We believe that the BDR penalty is due to an inconsistent downsampling process between matrix training and VTM simulation.

This contribution reports the simulation results by applying re-trained MIP matrix (MipSizeId=2). The downsampling function in MIP training has been modified to use the average of two samples from the original boundary. The simulation results have a luma BDR impact for AI/RA of -0.01%/-0.02% compared to VTM6.0 and decoding time impact for AI/RA of 100%/100%.

This is related to CE3-1, which was not retained. There was no need to present this.

References:
JVET-P0064
Decisions
This is related to CE3-1, which was not retained. There was no need to present this.
Citation