Graphics Reference
In-Depth Information
6.1 Real-time3Dvideoqualityevaluationstrategies
The measured image quality at the receiver-side can be used as feedback information to
update system parameters“onthefly”ina“QoE-aware”systemdesignapproach
[117][125].
However,measuring3Dvideoqualityinrealtimeisachallengemainlyduetothecomplex
nature of 3D video quality and also the fact that the amount of side-information to be
sent to measure the quality with RR methods is larger compared to 2D image/video
applications. The emerging RR and NR quality evaluation methods are based on image
features associated to the characteristics of the HVS. Some of these features are related
to image perception (e.g., luminance, contrast) and some are related to depth perception
(e.g., disparity, structural correlations). An appropriate selection of these features is crucial
to design an effective 3D image/video quality assessment method. The selected features
should be able to quantify image and depth perception related artifacts with a minimum
overhead. If the overhead is significant, the feasibility of deploying the designed RR
method is reduced. Figure 6.2 shows how the extracted edge information is employed to
measure 3D video quality in the RR method proposed in
[126].
In this method, lumin-
ance and contrast details of the original and distorted images are utilized to count for
conventional image artifacts, whereas edge information based structural correlation is em-
ployed to measure the structural/ disparity degradation of the 3D scene, which is directly
affecting rendering using colour plus depth map based 3D video. In order to reduce the
overheadforside-information(i.e.,extractedfeaturesofthereferenceimage)losslesscom-
pression mechanisms can be deployed for its compression. An extra effort should be also
made to send the side-information without corruption using a dedicated channel or highly
protectedforwardchannel.Visualattentionmodelscouldalsobeutilizedtofind3Dimage/
video features which attract significant attention during 3D viewing. However, a direct
relationship between visual attention and image perception for 3D images and video is yet
to be found. NR methods are the most suitable for real-time 3D video applications since
these do not consume any bandwidth for the transmission of side information. However,
their performance and application domain is limited since they rely solely on the received
3D image/video sequence and other contextual information (e.g., Hybrid-NR methods:
packet loss rate, bit-error rate). It may be impossible to count for all the artifacts imposed
along the end to end 3D video chain without referring to the original image sequence. This
is why most of the proposed NR metrics are limited to a specific set of artifacts
[127].
