Based on the proposed scheme (referred hereafter to as TB-mode), we develop three algorithms featuring different performance and complexity trade-offs. Ex-tensive experiments are carried out using the HEVC reference software (HM3.0) and the common test conditions [5] to compare their BD-rate savings relative to anchor encodings with TB-mode disabled. Rough estimates of their complexity characteristics are made by showing the encoding and decoding time ratios rel-ative to anchor settings. All three algorithms use parametric window functions.
The following summarizes these algorithms:
• Algorithm #1 applies TB-mode to 2N×2N prediction units (PUs)2. The vt is found by performing shape-adaptive template matching in a search range of MVp±8 pixels. For each 2N×2N PU, one flag is sent to switch adaptively between TB-mode and the usual inter mode. When the former is chosen, it codes two extra bits (at most) to specify the template shape (InvL, Rect-T or Rect-L). A separate set of window functions are designed for each distinct 2N×2N PU size.
• Algorithm #2 TB-mode with motion merging by signaling vtwith motion merging mechanism (See Section 4.3).
• Algorithm #3 simplifies Algo. #2 using two sets of simplified window functions. (See Section 4.4).
Table 4.1 presents the average BD-rate savings of these algorithms in different test classes and configurations. As can be seen, Algo. #1 has an average BD-rate
2CU, the basic compression unit as the MB in AVC, can have various sizes but is restricted to be a square shape. PUs are the various MB partitions having a square or rectangular shape with several sizes.
(Y) saving of 1.9% over all test cases, with a minimum of 1.2% and a maximum of 3.4%. Due to the extra computations needed for template matching, it doubles the decoding time while increasing the encoding time by about 70%.
Algo. #2 gives similar coding performance at a faster decoding speed than Algo. #1. Averagely, it perform an average BD-rate (Y) saving of 1.7% over all test cases, with a minimum of 0.9% and a maximum of 3.2%. with an average decoding time only 4% slightly longer than anchors’. In tests with High Efficiency configurations, its decoder can run even a bit faster. As expected, without having to perform template matching, its decoding complexity drops significantly. But.
somehow surprisingly, the vt signaled by motion merging seems to outperform that inferred by template matching. The reason may be twofold. First, the motion merging signaling mechanism incurs less overhead: while it needs only one extra bit to signal vt, Algo. #1 requires, on average, more than one bit to indicate the template shape. Second, template matching may result in poor vt due to coding noise. Recall that its search criterion is to minimize the error over the reconstructed pixels. Nevertheless, we believe Algo. #1 has plenty of room for further improvement.
Algo. #3 is the one with lowest computational complexity with an average BD-rate saving of 1.6% and a maximum saving up to 2.9%. This is attributed to the simplified window functions. But because this scheme utilizes PU-adaptive window selection from two sets of window functions, the encoder has to perform, for each PU, one extra motion search to evaluate TB-mode, which accounts for the slightly increased encoding time. But as for the decoder, the complexity is relative low due to a simple calculatioin of the predictor values.
Table 4.1: BD-rate savings and processing time ratios
Random Access High Efficiency Low Complexity
Algo. #1 #2 #3 #1 #2 #3
Class B -1.2 -0.9 -1.0 -1.4 -1.0 -0.9 Class C -2.0 -1.5 -1.6 -1.7 -1.2 -1.3 Class D -1.9 -1.6 -1.8 -1.6 -1.2 -1.4 All -1.7 -1.4 -1.2 -1.0 -1.1 -1.2 Enc. Time [%] 176 172 184 175 172 183 Dec. Time [%] 172 168 112 194 176 113 Low Delay High Efficiency Low Complexity
Algo. #1 #2 #3 #1 #2 #3
Class B -1.9 -1.5 -1.3 -2.4 -2.0 -2.0 Class C -2.3 -1.8 -1.7 -2.4 -1.9 -2.0 Class D -2.1 -2.0 -1.9 -2.2 -1.9 -2.0 Class E -3.3 -2.9 -1.8 -3.4 -3.2 -2.9 All -2.4 -2.1 -1.7 -2.6 -2.3 -2.2 Enc. Time [%] 157 150 162 155 149 159 Dec. Time [%] 220 155 109 269 155 110
4.6 Conclusion
Summarizing, in this chapter, we propose a bi-prediction scheme, which combines MVs found by template and block matchings with an optimized OBMC window function. Since the template MV is inferred on the decoder side, it has only to signal one block MV. This notion is further generalized by incorporating adaptive motion merging to allow a template-matching-free implementation. Three algo-rithms featuring different performance and complexity trade-offs are presented;
they all show moderate coding gains. The best of them produces an effect similar to performing partial motion merging for geometry/asymmetric motion partitions.
Chapter 5 Conslusion
In this dissertation, we have proposed an analytical perspective of MCP and devel-oped a generic scheme to reconstruct predictor from any irregular motion sampling structure. Proceeding in the generic reconstruction scheme, a bi-prediction com-bining TMP and BMC is introduced to further improve prediction efficiency. Most of techniques proposed in this dissertation are evaluated and cross-checked in CEs of the JCV-VC committee, showing very promising results. In the followings, we summarize the works in this dissertation and show how the proposed schemes can be further improved.
5.1 Motion-Compensated Prediction: An Ana-lytical Perspective
We found it convenient and insightful to interpret MCP as a motion sampling and reconstruction process. Such an interpretation reveals many important aspects of MCP which would otherwise be difficult to see. Using this notion, we provide a theoretical support for various MCP schemes from the sampling perspective. It is shown that the motion estimate found by template matching tends to be the motion associated with the template centroid and that TMP consistently out-performs SKIP prediction, but hardly competes with block motion compensation
(BMC) unless both the motion and intensity fields are less random or have high spatial correlation.