What are good CGS/MGS configurations for H.264 quality scalable coding?

摘要

Scalable video coding (SVC) encodes image sequences into a single bit stream that can be adapted to various network and terminal capabilities. The H.264/AVC standard includes three kinds of video scalability, spatial scalability, temporal scalability, and quality scalability. Among them, quality scalability refers to image sequences of the same spatio-temporal resolution but with different fidelity levels. Two options of quality scalability are adopted in H.264/AVC, namely CGS (coarse-grain quality scalable coding) and MGS (medium-grain quality scalability), and they may be used in combinations. A refinement layer in CGS is obtained by re-quantizing the (residual) texture signal with a smaller quantization step size (QP). Using the CGS alone, however, may incur notable PSNR penalty and high encoding complexity if numerous rate points are required. MGS partitions the transform coefficients of a CGS layer into several MGS sub-layers and distributes them in different NAL units. The use of MGS may increase the adaptation flexibility, improve the coding efficiency, and reduce the coding complexity. In this paper, we investigate the CGS/MGS configurations that lead to good performance. From extensive experiments using the JSVM (Joint Scalable Video Model), however, we find that MGS should be carefully employed. Although MGS always reduces the encoding complexity as compared to using CGS alone, its rate-distortion is unstable. While MGS typically provides better or comparable rate-distortion performance for the cases with eight rate points or more, some configurations may cause an unexpected PSNR drop with an increased bit rate. This anomaly is currently under investigation.