The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-328

“Rate-Distortion Based Dependent Coding for Stereo Images and Video: Disparity Estimation and Dependent Bit Allocation”

by Woontack Woo

December 1998

In this dissertation, novel coding schemes for stereo images/video are proposed. Recently, the demand for 3D imaging has been increasing because the stereoscopic method provides realism to 2D images. The price for this added realism is the doubling of data and thus, as in the single-channel case, the limited bandwidth of existing channels becomes the main bottleneck. To achieve an optimal coding gain for a pair of stereo images, we have proposed various efficient encoding schemes, which can be mainly grouped into two classes, blockwise dependent bit allocation and disparity estimation/compensation. In the proposed optimal blockwise dependent bit allocation scheme, the quantization parameters are selected simultaneously for blocks in both the reference image and the disparity compensated difference frame. In this manner, an average distortion measure can be minimized, while meeting any applicable bit budget constraints. In general, the bit allocation problem is complicated by the dependencies arising from using predictions based on the quantized reference image. Therefore, only approximate solutions are feasible in the case of motion compensated video. However, in the case of stereo images, an optimal solution can be estimated with reasonable complexity given the special characteristics of the ``binocular dependency.'' A fast algorithm is also proposed, which provides most of the gain at a fraction of the complexity. The proposed two hybrid estimation/compensation schemes are based on fixed and variable size blocks, respectively. The first scheme, modified overlapped block disparity compensation, can overcome drawbacks of conventional block-based schemes that use the smoothness constraints arising in causal neighborhoods by estimating a relatively smoother disparity field. Simultaneously, selective overlapped block disparity compensation for the blocks with higher prediction errors reduces blocking artifacts, while reducing computational complexity over the conventional overlapped matching scheme. The other scheme, quadtree-based hybrid block segmentation, can further improve the encoding efficiency along object boundaries. Similarly, a Markov Random Field model-based hierarchical approach allows the estimation of a consistent disparity field, even for small blocks. Furthermore, RD-based block segmentation and selective overlapped disparity compensation improve the encoding performance.

To download the report in PDF format click here: USC-SIPI-328.pdf (11.7Mb)