Evaluation of transform performance when using shape-adaptive partitioning in video coding

When combining non-rectangular (shape-adaptive) partitioning of inter pictures with a rectangular block transform, some of the transforms will be applied to a residual signal which originates from two different predictions. This is a condition that is never encountered in any of the existing video coding standards, and the impact of this on the performance of the transform coder has not been investigated. In this paper we investigate the effect of these mixed-signal blocks on the coding gain for various transforms, and we compare these against the optimal KLT gain. We find that, despite the transformed residual block being composed of two parts that were predicted from different areas of the reference picture, correlation within mixed blocks is very similar to that of normal blocks. The DCT is only marginally suboptimal w.r.t. KLT. KLT has practical issues that will reduce its coding gain or increase the signaling overhead: transform bases need to be quantized and transmitted, or a number of fixed bases needs to be chosen offline. Therefore we recommend DCT be used for all types of blocks.