Energy-aware scheme for the 3D-HEVC depth maps prediction

Abstract

This paper presents an energy-aware scheme to reduce the energy consumption on the 3D high-efficiency video coding (3D-HEVC) depth maps prediction. Besides, a qualitative discussion is presented for intra- and inter-frame predictions that conduced to the proposition of a simple energy-aware scheme. Through our analysis, the HEVC intra-prediction is applied over homogeneous regions, whereas bipartition modes are preferred to encode edge regions. Based on this fact, the Simplified Edge Detector (SED) is proposed to employ a fast intra-mode decision. The SED anticipates the blocks that are likely to be better predicted by the HEVC intra-prediction, avoiding evaluations of bipartition modes. On inter-prediction, the TZ Search (TZS) is employed in the 3D-HEVC reference software (HTM) to encode both texture frames and depth maps. However, considering the depth maps properties, lightweight fast algorithms should be considered instead of TZS. Thus, fast algorithms such as Diamond Search, Small Diamond Search (SDSP), and One-at-a-Time Search were evaluated in this paper, aiming to reduce the complexity and energy, whereas sustaining good coding efficiency. By analyzing the depth channel, this scheme (considering intra- and inter-predictions) is able to provide an encoding time reduction of 21.2---23.1 %. As drawback, the combined solution increases the BD-rate in 0.62---0.87 %, in the synthesized views. When considering general-purpose processors, our solution is capable of providing a reduction in the energy consumption ranging between 9.85 and 10.41 %, according to our software analysis using the running average power limit. By using the SDSP combined of SED algorithm instead of HTM-10.2 baseline solution, it is possible to achieve a reduction of about 54 % in the energy consumption, and about 1.8 times in the power dissipation, when running on a dedicated hardware design. Considering that depth maps are only used for view synthesis, a subjective quality assessment was performed using synthesized views, and the results demonstrate that our solution presents minimum quality losses.