Capture and Processing - 3DTV: Processing and Transmission of 3D Video Signals - page 60

and the approximation of a global 2-dimensional smoothness constraint by
combining multiple 1-dimensional constraints. The method can be divided
into three steps:
1. Block matching cost calculation : The aim of this step is to compute the
luminance difference between two searched areas in the left and right
images to find the corresponding pixels. In order to find them, the left
image (or right image) is divided into blocks and each block is compared
with a same size block in the right image (or left image) that is moved in a
defined range (called ''Horopter''), on the same line. A matching function
computes the cost associated with each tested area. The cost is calculated
as the absolute minimum difference of luminance intensities between the
compared blocks. Several parameters are adjustable during this step (e.g.
the size of the matching window, the size of the blocks, etc.) that can affect
the accuracy of the cost calculation. All the cost values throughout the
image pair are saved to be used in the next step.
2. Cost aggregation : The basic block matching cost calculation can yield
ambiguous values and incorrect matches can produce a lower cost than
the correct matches, due to a poor awareness of image content. Therefore,
a further constraint is added in the second step that enforces a smooth-
ness criterion by penalizing disparity value variations in the vicinity of
neighbourhood. Thus, the cost aggregation is based on a penalty-reward
scheme, and it progresses across eight directions originating from the
block under concern.
3. Disparity computation : Following the cost aggregation stage in the previ-
ous step, the disparity map is computed by selecting for each pixel p the
disparity d that corresponds to the minimum calculated cost. To avoid
problems arising in the image regions with poor texture information, a sec-
ondary parameter called the ''uniqueness constraint'' is introduced. The
incorporation of this parameter enforces the checking of the consistency
across a region and adapting the estimated disparity accordingly.
The MUSCADE Project [6] (as previously mentioned in Section 2.2) has
also dealt with the extraction of dense disparity information for four cameras
in the multi-view rig, following the rectification and colour correction steps.
The basis of the extraction of dense disparity information for a multi-camera
set is pair-wise extraction of stereoscopic correspondences.
The stereo processing stage is based on using a matching algorithm called
the Hybrid Recursive Matching (HRM) that is followed by an adaptive cross-
trilateral median filtering (ACTMF) post-processing step to regularize the
extracted disparity values. The initial values for the disparity are determined
by applying the hybrid recursive block matching method that operates both
in spatial and temporal (i.e. across successive video frames) directions to
yield smooth and consistent estimates. The initial disparity value estimation
process is applied in both directions, i.e. from right to left and from left