Digital Signal Processing Reference
In-Depth Information
(
)
For a non-occluded pixel
Pxy
in the synthesized image, its source may locates
,
v
v
only in the left image
(occluded in right view), or only in the right image
I
l
r
I
(occluded in the left view), or in both images (can be seen both in the left view and
the right image). So it should satisfy at least one part of eq.(1). Otherwise, it means
that the point is not visible in both views, and will be marked as a hole directly.
Please note that the mapping described in eq.(1) is not a one-to-one mapping.
Pixels that satisfying eq.(1) are regarded as candidates. We need to choose the correct
(nearest) pixel from all the possible candidates. The criterion is to select the right
pixel (either in left view or in right view) with the smallest depth value (or largest
disparity value if disparity maps are used).
3
Artifacts Detection
Because of the hardness of depth estimation, the depth map may inevitably contain
errors. When rendering based on erroneous depth maps, the resulting novel view will
contain artifacts. An example is given in Fig.2, in which, (a) is the left image, (b) is
the erroneous disparity map, and (c) is the initial novel view image. In the results,
obvious artifacts (e.g., holes, wrong colors and edges) can easily be found.
In order to detect these artifacts, we build a confidence map for every pixel in the
virtual view image according to the method proposed by Andrew J. Woods[8]. The
idea of the method is that the novel view should have similar pixel intensities,
gradients and Laplacians with the original image on the corresponding position. If
none of similar region exists in either of the two original images, then this pixel will
be given a low confident value. The corresponding confident map is shown in
Fig.2(d). Pixels with low confident value below a given threshold are regarded as
artifacts and marked as “Holes”, which should be filled up to produce a good result.
(a) Left image (b) Disparity map
Fig. 2.
Example of artifact detection
