In this dissertation, we introduced a parametric solution for OBMC (termed POBMC) and studied its properties from both theoretical and empirical aspects.
In the course, we found it convenient to approximate block MVs as motion samples taken at block centers; this expedient helped us to conceptualize the combination of OBMC with variable block-size motion partitioning. Because in practice it is far from adequate to describe the location of a block MV as a deterministic vari-able, we amended our solution to reveal its random nature. The novelty of our scheme was shown to require only the geometric relations between the prediction pixel and its nearby block centers for computing the window function. It thus lends itself to reconstructing temporal predictors from any sparsely and irregu-larly sampled motion data. The superiority of our scheme over similar previous works was confirmed by extensive experiments. Moreover, its performance was shown to be comparable to EAIF and QALF. Above all, it can work with either (or all) of them to yield an improvement that is nearly the sum of their separate effects.
Along with the coding tools in KTA2.4r1, part of this work [8] was submitted, for subjective viewing tests, in response to the HEVC Call for Proposals issued jointly by MPEG and VCEG in April, 2010 [2]. It was ranked 12 overall and 10 in low delay configurations among 27 proposals, in terms of the average mean opinion score [3]. The notion of POBMC has recently been extended to develop a reduced-overhead bi-prediction scheme [15][17]. Owing to its promising results, the technique is currently being evaluated in a core experiment of JCT-VC.
Chapter 4
Bi-Prediction Combining TMP and BMC
4.1 Introduction
Using the data accessible to the decoder for motion inference has recently emerged as a promising technique for next generation video coding. Template matching prediction (TMP) is a typical example of utilizing the decoder-side information for motion inference. It estimates the motion vector (MV) for a target block on the decoder side by minimizing the matching error over the reconstructed pixels in its immediate inverse-L-shaped neighborhood (usually termed the template).
Motion merging [28] and OBMC [19] also follow the same rationale. They both view the received motion data as a source of information about the motion field and forms a better prediction of a pixels intensity based on its own and/or nearby block MVs.
For the above reasons, we are led to develop a biprediction scheme, which incurs only the overhead for unidirectional prediction. The idea is to combine MVs resulting from the template and block matchings with OBMC. Of particular interest in this combination is that the template MV is inferred on the decoder side;
it thus has only to signal one block MV while attaining biprediction performance.
The choice of a proper window function is critical in these applications. We
approach this problem through the parametric OBMC framework in Chapter 3.
The resulting window function is shown to resemble a particular type of geometry motion partitioning as shown in Fig. 4.4(c) with its MVs located on the diago-nal running from the upper left to the lower right. Particularly, asymmetric-like partitioning as shown in Fig. 4.4(a),(b) arises when the template region is of rectangular shape and is located to the left or on the top of a target block.
To gain more insights into this biprediction scheme, Fig. 4.1 plots the mean-square prediction error surface with the single use of TMP MV for motion com-pensation of the target block. It is seen that this MV tends to minimize the prediction error in the upper left quarter, a result that is intuitively agreeable since it approximates the true motion associated with the template centroid. This observation implies that the block MV should be managed to contribute more to minimizing the error in the remaining part, especially in the bottom right region.
We thus also propose in this chapter a new search criterion for the block MV to achieve this objective.
The proposed bipredictioin scheme can be further generalized to allow a template-matching-free implementation, which replaces the template MV with a decoded MV specified by a mechanism similar to motion merging. The approach is to replace the template MV with one of those for prediction units (PUs) in a causal neighborhood of the current PU. In particular, the selection of MV is made adaptive by adopting the same signaling mechanism as for motion merging [28].
Depending on the direction in which the MV (and other motion parameters) is derived, a separate window function is applied for OBMC. Remarkably, this scheme can produce, with less overhead, an effect similar to combining asymmet-ric/geometry motion partitions [14][18] and PU-based motion merging [28].
Figure 4.1: Mean-square prediction error surface of TMP using the sequence
”Football”.
In the following sections, we will first detail the concept of the joint application of TMP and BMC. A low complixity and template-matching-free implementation is then introduced. We will also show the experiments conducted using the HEVC reference software and the common test configurations to evaluate three variants of this scheme. The best of them achieves an average BD-rate saving of 2.2%, with a minimum of 0.2% and a maximum of 4.7%. It is observed that the average decoding time increases by 10% while the encoding time doubles.