This thesis addresses the problem of computing 2-D disparity fields from stereo image pairs and applying them to intermediate view reconstruction (IVR) via disparity-compensated interpolation. Intermediate views are calculated using a linear filter with angle-dependent coefficients. Two existing disparity estimation algorithms are adapted to perform IVR, and results are used for parallax adjustment of still stereo images. In one case, the well-known block matching (BM) technique is used, and several novel algorithm enhancements are proposed. The final BM scheme employs three-component estimation, a spatial smoothness constraint, and a quadtree structure. A procedure for targeting problematic blocks that requires splitting based on robust estimation is proposed, and an efficient approach for the reestimation of sub-blocks is developed. A technique for eliminating component-mismatches in stereo pairs is also examined. The accuracy of estimations based on these "balanced" images is seen to increase.;In the other case, the ill-posed problem of obtaining dense disparity maps is addressed, and the method of regularization is used to compute pixel-based vector fields for intermediate views. The conclusion is that although single image reconstruction results are comparable in both cases, the approach based on regularization is superior to the block-based scheme for a dynamic sequence of such reconstructions. Both approaches are applied to stereo parallax adjustment for still images, and numerous experimental results are included.
展开▼
