Game Development Reference
In-Depth Information
11.3 3D Quality Assessment
In 3D quality assessment, there are eight databases that are commonly recog-
nized in the research community, which include the LIVE 3D Image Quality Data-
base Phase I (Moorthy et al.
2013
), LIVE 3D Image Quality Database Phase II
(Chen et al.
2013
), IRCCyN/IVC 3D Images Database (Benoit et al.
2008
), MICT
3D Image Quality Evaluation Database (Sazzad et al.
2009
), Ningbo University 3D
Image Quality Assessment Database-Phase I and II (Wang et al.
2009
; Zhou et al.
2011
), Tianjin University 3D Image Quality Assessment Database (Yang et al.
2009
)
and MMSPG 3D Image Quality Assessment Database (Goldmann et al.
2010
).
The LIVE 3D Image Quality Database I and II contain distortion types including
JPEG 2000, JPEG, white noise, Gaussian blur and fast fading, in which both sym-
metric and asymmetric cases are considered. IRCCyN/IVC 3D Images Database
contains 96 images with distortion types like JPEG2000, JPEG, and gaussian blur.
InMICT 3D Image Quality Evaluation Database, only JPEG distortion is considered,
in which both asymmetrically and symmetrically distorted images are evaluated. The
Ningbo University 3D Image Quality Assessment Database I only contains asym-
metrically distorted stereoscopic images, and Ningbo University Database Phase II
only contains symmetrically distorted stereoscopic images. Tianjin Database con-
tains JPEG2000, JPEG, and white noise distortions with size from 320
×
240 to 1,024
×
768. The MMSPG 3D Image Quality database contains 100 distorted images pairs
with different camera distances.
In the literature, existing 3D IQAs can be categorized into two groups. The first
type is built on successful 2D-IQAmethods. Depending whether including the depth
or disparity information, these approaches can be further divided into two subcat-
egories. In the first subcategory, the depth information is not explicitly used. For
example, in Benoit et al. (
2008
), Brandão and Queluz (
2008
), 2D-IQA measures
including PSNR, SSIM and video quality metric (VQM) were employed to the left-
and right-view images of 3D videos separately and then combined to a 3D quality
score. In Campisi et al. (
2007
), four 2D-IQA metrics, namely SSIM, universal qual-
ity index (UQI), C4 and RR-IQA as well as three approaches, termed as average
approach, main eye approach, and visual acuity approach, were applied with an aim
of measuring the perceptual quality of stereoscopic images. In the second subcate-
gory, the depth information is incorporated with 2D-IQA to derive the quality. The
depth map is estimated firstly from the left and right views. Then all the scores are
combined to the final 3D score. Moreover, ten 2D-IQAmetrics including the popular
PSNR, SSIM, multiscale SSIM (MS-SSIM), VIF, etc., are used to evaluate the 3D
image quality. Yang et al. (
2009
,
2010
) proposed a 3D IQA algorithm based on the
average PSNR of two views and the absolute difference of the disparity map.
The second group 3D-IQA or 3D-VQA approaches focus on establishing the 3D
quality models directly without referring to the 2D IQA scores. For example, in
Gorley and Holliman (
2008
) SIFT and RANSAC are used to perform matching on
stereo images, and Gorley et al. predicted the stereoscopic image quality by feature
mapping between them. The depth perception is considered in Zhu andWang (
2009
),
and a multichannel vision model based on 3D wavelet decomposition is proposed to
access the stereoscopic video quality.
