Graphics Reference
In-Depth Information
These objective measures may or may not strongly correlate with the quality attributes
of 3D video as measured by subjective tests. Studies have found out that there is a high
correlation between subjective ratingsandindividualobjectivequality ratingsof3Dvideo
components (e.g., average PSNR and SSIM of left and right video or colour and depth
video)
[123].
For instance, depth perception is highly correlated to the average PSNR
of the rendered left and right image sequences
[123].
This could be due to the loss of
correspondence between left and right objects and reduction of monocular depth cues as
a result of compression and transmission errors. This means that we could use individual
objective quality measures of different 3D video components to predict the true user per-
ception in place of subjective quality evaluation, through a suitable approximation derived
based on correlation analysis. However, with some 3D source representations such as the
colour and depth map 3D image format, it may be difficult to derive a direct relationship
between objective measures and subjective quality ratings. For instance, the objective
quality of the depth map may have a very weak correlation on its own with the overall
subjective quality, because the depth map is used for projecting the corresponding col-
our image into 3D coordinates and it is not directly viewed by the end users. Individual
quality ratings of left and right views may not always account for depth reproduction of
the scene. Therefore, the next phase of 3D objective quality metrics includes a method-
ology to quantify the effect of binocular disparity of 3D scenes in addition to a conven-
tional image/video quality assessment methodology. For instance in
[122],
in addition to
image quality artifacts, disparity distortion measures were also incorporated to evaluate
the overall 3D video quality. The article showed improved performance over the method
which does not account for the correspondence information of stereoscopic views. The
latest 3D image/video quality metrics evaluate depth reproduction in addition to usual
image artifacts (such as blockiness) using specificimagefeatures(e.g.,edge,disparityand
structural information of stereoscopic images) which are important for the HVS in both
2D and 3D viewing. For instance the method proposed in
[124]
shows high correlation
values with subjective quality results (Mean Opinion Score, MOS): the correlation coeffi-
cient with subjective quality ratings is as high as 0.95; this outperforms the method based
on 2D image quality + disparity
[122]
and other conventional 2D quality metrics separ-
ately applied to left and right views (see Table 6.1). The reported performance figures in
Table6.1areobtainedusingthesame3Ddataset.Theseobservationsconfirmthataccurate
3D image quality metrics should be designed to also consider binocular disparity distor-
tions. All the methods described above are Full-Reference (FR) methods and need the
original 3D image sequence to measure the quality by comparison, hence they arenotsuit-
ablefortheevaluationofthequality“onthefly”inreal-timetransmissionapplicationssuch
