Correlation noise modeling for multiview transform domain Wyner-Ziv video coding

Abstract

Multiview Wyner-Ziv (MV-WZ) video coding rate-distortion (RD) performance is highly influenced by the adopted correlation noise model (CNM). In the related literature, the statistics of the correlation noise between the original frame and the side information (SI), typically resulting from the fusion of temporally and inter-view created SIs, is modelled by a Laplacian distribution. In most cases, the Laplacian CNM parameter is estimated using an offline approach, assuming that either the SI is available at the encoder or the originals are available at the decoder which is not realistic. In this context, this paper proposes the first practical, online CNM solution for a multiview transform domain WZ (MV-TDWZ) video codec. The online estimation of the Laplacian CNM parameter is performed at the decoder based on metrics exploring both the temporal and inter-view correlations with two levels of granularity, notably transform band and transform coefficient. The results obtained show that better RD performance is achieved for the finest granularity level since the inter-view, temporal and spatial correlations are exploited with the highest adaptation.