Video quality testing methods

Comparing subjective video quality testing methodologies
M. Pinson and S. Wolf
Institute for Telecommunication Sciences (ITS), National Telecommunications and Information
Administration (NTIA), U.S. Department of Commerce
ABSTRACT severity and ordering of impairments within the test
session. With DSCQS, context effects are minimized
International recommendations for subjective video since viewers are shown pairs of video sequences (the
quality assessment (e.g., ITU-R BT.500-11) include reference sequence and the impaired sequence) in a
specifications for how to perform many different types of randomized order. Viewers are shown each pair twice.
subjective tests. Some of these test methods are double After the second showing, viewers are asked to rate the
stimulus where viewers rate the quality or change in quality of each sequence in the pair. The difference
quality between two video streams (reference and between these two scores is then used to quantify changes
impaired). Others are single stimulus where viewers rate in quality. The resulting scores are not significantly
the quality of just one video stream (the impaired). Two impacted by memory-based biases from previously viewed
examples of the former are the double stimulus continuous video sequences (see section 6.3.2 of [1]). Since standard
quality scale (DSCQS) and double stimulus comparison double stimulus methods like DSCQS1 provide only a
scale (DSCS). An example of the latter is single stimulus single quality score for a given video sequence, where a
continuous quality evaluation (SSCQE). Each subjective typical video sequence might be 10 seconds long,
test methodology has claimed advantages. For instance, questions have been raised as to the applicability of these
the DSCQS method is claimed to be less sensitive to testing methods for evaluating the performance of
context (i.e., subjective ratings are less influenced by the objective real-time video quality monitoring systems.
severity and ordering of the impairments within the test By contrast, single stimulus continuous quality evaluation
session). The SSCQE method is claimed to yield more (SSCQE) allows viewers to dynamically rate the quality of
representative quality estimates for quality monitoring an arbitrarily long video sequence using a slider
applications. This paper considers data from six different mechanism with an associated quality scale. This
subjective video quality experiments, originally performed relatively new method provides a means for increasing the
with SSCQE, DSCQS and DSCS methodologies. A sampling rate of the subjective quality ratings. Having
subset of video clips from each of these six experiments subjective scores at a higher sampling rate would be
were combined and rated in a secondary SSCQE useful for tracking rapid changes in quality and thus would
subjective video quality test. We give a method for post- be more useful for evaluating real-time quality monitoring
processing the secondary SSCQE data to produce quality systems. Proponents of the SSCQE methodology argue
scores that are highly correlated to the original DSCQS that it can be used to assess widely time-varying quality of
and DSCS data. We also provide evidence that human long video sequences in a way that DSCQS cannot (see
memory effects for time-varying quality estimation seem sections 6.3 and 6.4 of [1]). However, questions have
to be limited to about 15 seconds. been raised regarding the accuracy of SSCQE when
Keywords: single stimulus continuous quality evaluation compared to DSCQS. Since viewers only see and rate the
(SSCQE), double stimulus continuous quality scale quality of a single video stream (i.e., single stimulus with
(DSCQS), double stimulus comparison scale (DSCS), no immediate reference picture), contextual effects may be
correlation, video quality, image quality, subjective present. Individual viewers’ scores might also drift over
testing, picture quality. the course of the test (e.g., viewers might concentrate on

1. INTRODUCTION 1
Reference [1] presents a new form of double stimulus
The double stimulus continuous quality scale (DSCQS) testing called the simultaneous double stimulus for continuous
method of performing subjective tests is widely accepted evaluation (SDSCE) that utilizes side-by-side presentation of the
as an accurate test method with little sensitivity to context original and impaired clips rather than randomized time
effects (see Appendix 3 to Annex 1 in [1]). Context ordering. However, this method has the drawback that the
effects occur when subjective ratings are influenced by the viewer must shift attention between the right and left
presentations.

moving the slider in the proper direction to track changes One hundred is added to the difference, shifting the range
in quality and hence loose track of the absolute slider to [0, 200]. Here, 0 is the worst quality, 100 is the same
position on the rating scale), adversely impacting the quality as the reference, and values greater than 100
method’s reliability. Differences in viewers’ reaction indicate quality better than the reference. Scores greater
times to quality changes would also reduce SSCQE’s than 100 seem to be limited to the first several seconds of
accuracy. the video scene (i.e., viewers seem to require about 6 to 8
seconds to move the slider to the proper position after a
This paper presents results from a recent subjective meta-
scene transition from a low quality scene to a reference
experiment where video clips from six separate subjective
high quality scene). The hope was that SSCQE with
tests were combined and evaluated using an SSCQE
hidden reference removal could provide time varying
experiment. Throughout this paper ‘original ratings’ will
quality assessments with the added benefits of double
refer to the subjective ratings associated with the original
stimulus testing. SSCQE with hidden reference removal
subjective experiment. ‘Secondary ratings’ will refer to
has also been proposed by the Video Quality Experts
the subjective ratings associated with the SSCQE meta-
Group (VQEG) for future tests of reduced reference - no
experiment performed on those same video clips. Five of
reference (RRNR) objective video quality monitoring
the original experiments were some form of double
systems [2]. Viewer training instructions for our SSCQE
stimulus testing while one of the original experiments was
experiments were taken directly from [2].
itself an SSCQE experiment. By comparing the secondary
mean opinion scores (MOSs) from the SSCQE meta- Two original data sets utilized DSCQS which is described
experiment to the original MOSs from the six original in section 5 of [1]. One study indicates that DSCQS is not
experiments, we have begun to find answers to the influenced by contextual effects (see Appendix 3 to Annex
questions that have been raised. Do SSCQE and DSCQS 1, in [1]). However, a problem we discovered with
yield inherently different subjective scores? Are DSCQS testing is that viewers occasionally switch scores
contextual effects and the lack of an immediate reference (e.g., enter their opinion of the reference video where the
picture problematic for SSCQE? To what extent does impaired video score should be recorded, and vice versa.)
previously viewed video impact the viewer’s current This can be demonstrated by an examination of individual
opinion score in SSCQE? viewer scores for one video sequence from subjective
laboratory seven for the VQEG 525-line low quality full-
2. SUBJECTIVE TESTING reference test [3]. In Figure 1, most individual viewer
This section describes the three subjective video testing scores are scattered in a rough distribution around 77.
methodologies that were used, the original subjective The subjective response of -75 is indicative of viewer
experiments, and the meta-experiment that was used to switching. Even though we have just plotted results for
obtain the secondary subjective ratings. Throughout this one video clip, viewer score switching seems to be a fairly
paper, a ‘data set’ will refer to one subjective video common occurrence.
quality experiment.

2.1 SSCQE, DSCQS, and DSCS
One original data set and the secondary data set utilized
SSCQE. Our SSCQE experiments used hidden reference
removal, a data post-processing step that is not described
in [1].2 With hidden reference removal, the reference
video sequences are presented during the test session, but
viewers are not aware that they are evaluating the
reference video. The viewer’s opinion of the reference
video sequence is subtracted from the viewer’s opinion of
the impaired video sequence. Since each opinion is in the
range [0, 100] the difference is in the range of [-100, 100].

2
SSCQE with hidden reference removal was first described
to the authors of this paper by Philip J. Corriveau at
Communications Research Centre (CRC), Ottawa, Ontario, Figure 1. Histogram of individual DSCQS scores for one video
Canada. sequence.

Three of the original subjective data sets employed a third limited range of quality than data set six. The VQEG data
testing methodology known as the double stimulus sets have more viewers than the other data sets since four
comparison scale (DSCS). DSCS is described in section independent international laboratories were used.
6.2.4.1 ITU-R Recommendation BT.500 [1]. DSCS
Data set twelve was generated after the publication of [4].
viewers are presented with a pair of video sequences.
Data set twelve was also performed by NTIA/ITS using
Like the DSCQS method, the order within the pair is
SSCQE, modified for hidden reference removal as
randomized, but unlike DSCQS, the pair is presented once
described earlier. Data set twelve used ten 1-minute video
instead of twice. With DSCS, the viewers directly rate the
sequences and four video systems. The video sequences
difference between the first and second video sequence on
were split into two viewing sessions, with two random
a discrete seven point scale (as opposed to DSCQS where
orderings for each session. Each viewer watched one
both the first and second video sequences are each rated
session only, to minimize viewer fatigue. All viewers
on a continuous quality scale). The viewers indicate
rated the reference video sequence as a hidden reference.
whether the video quality of the second clip was better,
worse, or the same as the first clip. The DSCS
Table 1 Original Subjective Tests
experiments used in this study were well-balanced with
respect to order of presentation and range of video quality. Data Method Viewers Video ITU-R
Set Clips BT.500-
In summary, DSCQS viewers make two absolute ratings
on a continuous scale at discrete times; DSCS viewers One DSCS 32 42 BT.500-3
make one difference rating on a discrete scale at discrete
times; and SSCQE viewers make absolute ratings on a Two DSCS 32 105 BT.500-3
continuous scale continuously over time. Three DSCS 32 112 BT.500-3
2.2 Original subjective experiments Four DSCQS 67 90 BT.500-8
Video sequences for the SSCQE meta-experiment were Six DSCQS 80 90 BT.500-8
taken from six different video quality tests. These six
original experiments examined impairments from Twelve SSCQE 32 40 BT.500-
television systems, with a particular emphasis on MPEG-2 10
coding impairments. Some of the video systems included
MPEG-1 coding, VHS record/playback, multiple- 2.3 Secondary subjective experiment
generation dubbing with 1/2 inch professional tape The secondary subjective test was performed by
recorders, and MPEG-2 bit-streams corrupted with digital NTIA/ITS using SSCQE with hidden reference removal.
errors. Table 1 summarizes each of the six video tests, Video clips were chosen from the six original data sets.
listing the subjective method used, the number of viewers, This subjective meta-experiment was designed to last 30
the total number of video clips evaluated, and the version minutes. A panel of 20 viewers was split into two groups,
of ITU-R Recommendation BT.500 that was used.3 with each group seeing one of the two possible orderings.
Data sets one, two, three, four, and six are described in To limit the duration of the secondary test, five scenes and
detail in [4], where the data set numbers shown in the four video systems were chosen from each of the original
table correspond to those that were used in [4]. As a brief subjective data sets one, two, three, four, and twelve (for
summary, subjective experiments one, two, and three were data set twelve, five 9 second scene segments were
performed by NTIA/ITS and used scenes that were 9 selected from the 1-minute scenes). An indicator for the
seconds long. Subjective experiments four and six were coding difficulty of each original scene was obtained by
performed by VQEG during their first round of full- averaging the subjective MOSs for that scene across all
reference television (FRTV) tests and used scenes that video systems. Scenes were then chosen to evenly span
were 8 seconds long [3]. Data sets four and six have some the full range of available quality in each data set. Video
overlap, with all source scenes and two video systems systems were chosen in a similar manner.
being in common to both data sets. Data set four included
only high-quality television systems with a much more The entirety of data set six was included in the secondary
subjective test due to its high accuracy (i.e., large number
of viewers), wide quality distribution, and because we
3
We used the numbering established in [4] so that additional wished to explicitly compare DSCQS quality scores with
information on each data set could be located within that SSCQE quality scores. We will thus present a more
reference easily and accurately.

detailed quality comparison analysis for data set six. Also at the end of the previous super-scene to easy-to-code
because of the overlap between data sets six and four, clips at the start of the current super-scene. The gradual
more video clips were included from data set four than fall off is due to increasing coding difficulty within the
from data sets one, two, three, and twelve. current super-scene. The above effect is more pronounced
for the curve that has data set twelve removed.
To allow time for SSCQE scores to stabilize for hidden
reference removal and to simplify analysis, the five scenes
2.4 Putting scores on a common scale
from each data set were treated as a 45-second super-
scene. To minimize the magnitude of slider adjustments Although all six original subjective tests conformed to
required within the super-scene, original clips were some version of Rec. ITU-R Recommendation BT.500,
ordered from easy-to-encode to hard-to-encode. For data each test had its own unique set of variables that
sets four and six, which have 8 second video clips, an influenced the subjective scores (e.g., subjective scale,
extra 5 seconds of video was inserted at the beginning of viewer expectations, and range of video quality included
the super-scene to fill out the 45 seconds (the scores from in the test). Thus, even when two subjective tests utilize
these extra 5 seconds were not intended to be analyzed). exactly the same subjective testing methodology and
Test presentation ordering was randomized over super- scale, there is normally some gain and shift of one set of
scenes and video systems, with the added constraint that subjective scores with respect to the other. Having the
the same super-scene or video system would never appear secondary subjective test ratings allowed us to linearly
twice in a row. scale and shift each of the original data sets (using linear
regression techniques) to one common scale, namely, the
Figure 2 plots the average viewer time response for a secondary subjective scale. This simplified our analysis
super-scene from the secondary experiment (i.e., averaged and data set comparisons without affecting computation of
across all viewers, super-scenes, and video systems under such things as Pearson correlation coefficients.
test). Two curves are shown on the plot, one including all
of the data and the other excluding data set twelve, since For the linear regression (and all other analyses), each
the clips within data set twelve’s super-scene were original data set was considered separately. In other
mistakenly ordered from most difficult to encode to words, each original data set had its own unique scaling
easiest to encode. and shifting factors that were used to map the original
ratings to the secondary subjective scale. Thus, the 20
original clips that were in common to data sets four and
six will appear twice on the secondary scale, once from
data set four and once from data set six. Likewise, the
analysis to be presented includes 205 video clips even
though the viewers rated only 185 video clips in the
secondary subjective test.

3. REPRODUCIBILITY OF SUBJECTIVE
RATINGS
The SSCQE method produces a subjective rating every
half second, whereas the DSCQS and DSCS methods
produce one subjective rating for each 8-10 second long
video clip. There has been work on methods to convert
SSCQE subjective ratings to DSCQS subjective ratings.
One study examined both a 10 second averaging process
of the SSCQE ratings and a non-linear averaging process,
Figure 2. Score within super scene, averaged across viewers, yielding promising results [5]. Another study found that
super-scenes, and systems under test. the DSCQS subjective rating of a 30 second video clip is
different than the average SSCQE subjective rating for
Notice that the ordering of scenes within the super-scene
those 30 seconds [6].
results (on average) in a sharp increase in viewer score
during the first seven seconds and a gradual fall off of Preliminary analysis we performed on subjective ratings
scores during the last 20 seconds. The sharp increase provided by the Communications Research Centre (CRC)
represents the sharp transition from the hard-to-code clips showed that the SSCQE rating located at the end of each

video clip is highly correlated with DSCQS scores of Table 2 compares the original and secondary subjective
those same video clips.4 This observation was verified for scores for each data set. The column labeled ‘Clips’
the subjective meta-experiment presented in this paper. contains the number of video clips held in common
Thus, unless otherwise specified, the SSCQE ratings at the between the two tests. The ‘Variance’ column is the total
end of the 8 or 9 second video clip are used for all variance of those clips within the original data set after
comparisons of secondary and original ratings. These scaling to match the secondary data’s 100 point scale.
comparisons include Pearson linear correlation The ‘Correlation’ and ‘RMSE’ columns provide
coefficients and root mean square error (RMSE). RMSE comparisons of the original and secondary subjective
values are calculated as the root mean square error of a scores. The bottom row labeled ‘All’ lists these values
linear regression, where the scaled original subjective data computed over the set of original video sequences selected
is the predictor variable (x-axis) and the secondary for inclusion into the secondary subjective test. The
subjective data is the response variable (y-axis). RMSE is correlation coefficients and RMSE values in Table 2 are
an approximation of the standard deviation of the impressive considering the different subjective test
residuals (or observed errors) when the scaled original methodologies that were used.5
subjective data is used to estimate the secondary
Data set four has a total variance significantly lower than
subjective data [7].
all the other data sets. Notice that data set four’s video
Because the total variance of each original data set is clips are clustered close together at the high quality end of
different (i.e., each original data set spans a different the scale in Figure 3, but the spread of points around the
range of video quality), correlation coefficients do not tell 45º line is comparable to that seen in other data sets. Data
the whole story. RMSE can complete the picture because set four’s correlation is significantly lower than the
it provides an estimate of the prediction errors with correlations for data sets one and two, but the RMSE
respect to a common video quality scale, namely the values for these three data sets are similar in magnitude.
secondary subjective data scale. This example demonstrates that simply looking at
correlation coefficients can be misleading.
3.1 Repeatability
For each data set, Figure 3 plots the secondary subjective Table 2 Comparison of Individual Data Sets
ratings versus the original subjective ratings (after scaling Data Set Clips Variance ρ RMSE
and shifting). A score of 100 corresponds to excellent
quality. One 20 238 0.961 4.55
Two 20 371 0.963 5.55
Three 20 325 0.888 9.60
Four 35 44 0.782 5.34
Six 90 170 0.917 5.70
Twelve 20 314 0.929 7.24
All 205 254 0.936 6.01

Data sets one, two, four, and six have RMSE values that
are reasonably similar. Data sets three and twelve have
substantially higher RMSE values. To understand these
higher RMSE values, we examine the effect of clip
presentation ordering of the original and secondary
Figure 3. Secondary versus original subjective ratings, plotted subjective tests. Because the results for data sets three
by data set.
5
For the secondary data, we also examined SSCQE without
4
This data was received by means of a private hidden reference removal. Here, the squared correlation
communication with Philip J. Corriveau at Communications coefficient (an estimate of the fraction of the variance explained)
Research Centre (CRC), Ottawa, Ontario, Canada. was reduced by 0.05 on average.

and twelve depend upon only 20 data points, single Close analysis of data set twelve identified one outlier.
outliers in these data sets can have a significant impact on For both orderings of the secondary data, this video clip’s
RMSE. Table 3 expands upon Table 2, listing results subjective ratings are noticeably shifted away from the
separately for each of the two secondary clip presentation main cluster of data. This data point is plotted in Figure 4,
orderings. Most of the correlations and RMSEs are at an original value of approximately 60 and a secondary
reassuringly similar for order #1 and order #2. value of approximately 80. This particular clip had a
sharp drop in quality during the last 2 seconds. For some
The largest RMSE overall belongs to data set three. The
reason, the original viewers seem to have reacted faster to
largest difference between order #1 and order #2
the rapid drop in quality than the secondary viewers. If
correlations and RMSEs likewise belongs to data set three.
this clip is removed from data set twelve, the correlation
Investigations revealed that most of data set three’s
between the original and secondary data rises from 0.929
differences seen in Table 3 are caused by a single outlier
to 0.962 and the RMSE drops from 7.24 to 5.32. While
clip in the secondary data. This clip occurred very close
the former RMSE (7.24) is somewhat higher than the
to the beginning of order #2. The video system from
usual (see Table 2), the latter RMSE (5.32) puts data set
which this clip came was a VHS tape recording. As studio
twelve into the range of data sets one, two, four, and six.
quality reference video had not yet been presented (except
With the six outliers from data sets three and twelve
during a training session), several viewers rated this VHS
removed, the secondary subjective data set can be
video clip as having perfect quality. By contrast, for order
replicated from the original data sets with an RMSE of 5.7
#1 this same video clip was presented later in the session
or less.
after viewers had a chance to see very high quality video.
For order #1, the resulting secondary subjective ratings
Table 4 Influence of VHS Outliers on Data Set Three
compare more favorably to the original subjective ratings.
The other four scenes that were presented as part of that Conditions Order 2 Order 2 Overall Overall
same super-scene exhibit the same ordering effect, ρ RMSE ρ RMSE
although to a lesser extent. Thus, multiple presentation
orderings appear to be important for good SSCQE All 20 clips 0.802 12.40 0.888 9.60
subjective test design. Ideally, each viewer should see a Discard 1 0.872 10.28 0.924 8.17
unique randomized order. Then, clip ordering effects are VHS
mitigated when averaging across viewers. outlier
Table 3 Impact of SSCQE Clip Presentation Ordering Discard all 0.960 6.36 0.986 3.92
5 VHS
Data Order 1 Order 2 Order 1 Order 2 outliers
Set RMSE RMSE ρ ρ
One 3.62 6.54 0.976 0.920
Two 6.21 5.42 0.951 0.966
Three 8.02 12.40 0.931 0.802
Four 5.06 6.53 0.749 0.762
Six 6.31 7.25 0.894 0.882
Twelve 8.02 7.30 0.917 0.926
All 6.31 7.54 0.932 0.901

Table 4 shows how the correlation rises and the RMSE
falls when the VHS outliers are removed from data set
three. Notice that, once all five VHS outliers have been
removed, the correlation and RMSE of the data set as a
whole (‘Overall ρ’ and ‘Overall RMSE’) are better than
Figure 4. Secondary versus original subjective data set twelve,
those seen in Table 2 for the other data sets.
order #1 and order #2, plotted separately.

3.2 Lab to lab comparisons, data set six • The SSCQE rating at the end of the 8-9 second video
Data set six was collected using four different testing clip is used (i.e., an 8-9 second stabilization period
laboratories, referred to in [3] as labs 2, 3, 5, and 7. Table before sampling the SSCQE rating).
5 presents the cross correlations between the original labs
and the secondary SSCQE data for data set six. The lab- 4. MEMORY EFFECTS
to-lab correlations clearly show that lab 3 was an outlier. One interesting question to examine is the impact of
The reasons for this have not been documented. human memory on SSCQE results. To what extent is the
Correlations between SSCQE and labs 2, 5, and 7 (i.e., viewer’s current impression of the video quality dependent
ranging from 0.902 to 0.929) are comparable to the upon past impairments? There is some evidence that
correlations between labs 2, 5, and 7 (ranging from 0.913 viewers have non-symmetrical memory in that they are
to 0.935). These results indicate that the secondary quick to criticize degradations in video quality but slow to
SSCQE subjective ratings are nearly comparable to the reward improvements [5].
original DSCQS subjective ratings from each of these
three labs. The impact of specific training instructions on the
subject’s memory (for quality estimation purposes) has not
Table 5 Cross Correlations Between Original Labs and been extensively studied. It is reasonable to assume that
Secondary SSCQE Ratings for Data Set Six these instructions (or lack thereof) could influence results
of any investigation into SSCQE memory effects. For
Lab2 Lab3 Lab5 Lab7 both data set twelve and the secondary experiment, our
viewers were encouraged (in the training session) to
SSCQE 0.911 0.682 0.902 0.929 respond to rapid changes in quality (see training
Lab7 0.933 0.727 0.935 instructions in [2]). Therefore, the results presented here
might be different from what other researchers have
Lab5 0.913 0.807 found. We examine three approaches to assess the impact
of memory effects on subjective results. These
Lab3 0.747
approaches are based on the repeatability of subjective
scores and the importance of scene length.6
Table 6 lists the number of viewers (N) and confidence
4.1 The nine second approach
intervals (CI) for the original labs taken individually and
altogether, and the secondary SSCQE data. For roughly The secondary SSCQE experiment used 8 to 9 second
the same number of viewers (i.e., 20), the CI for the scene segments drawn from the original data sets in Table
secondary data appears to be comparable to the original 2. If memory effects extended much beyond 9 seconds,
data CI for one lab. we would expect to see a significant impact on the
correlation coefficients and RMSE values that are
Table 6 Lab-to-Lab Comparison of Average Confidence presented in Table 2. In other words, if memory extended
Interval significantly beyond 9 seconds, why would instantaneous
SSCQE data sampled at the end of each 9 second scene
Lab2 Lab3 Lab5 Lab7 DSCQS SSCQE agree with the original subjective data where viewers only
N 17 18 27 18 80 20 saw that 9 second scene? Thus, Table 2 by itself presents
an argument that significant memory effects do not extend
CI 8.12 10.19 5.27 8.19 4.05 5.94 much beyond 9 seconds for our tests.
It is possible that some viewers may interpret scene
3.3 Summary changes as cues to update their opinions, thus artificially
The results presented here indicate that SSCQE may be truncating their natural memory effect to 9 seconds. Let
used to approximate DSCQS and DSCS subjective ratings us now examine an approach that utilizes longer video
provided: sequences.
• SSCQE data is processed using the hidden reference
removal technique.
6
The subjective testing community has debated the validity
• At least two randomized viewer orderings (i.e., of using short program segments for SSCQE testing. For
sessions) are performed in the SSCQE experiment. instance, section 6.3.1 of [1] recommends using program
segments at least 5 minutes long.

4.2 Correlating SSCQE samples from two used only the previous 7 or 8 seconds of video for forming
experiments their SSCQE quality decision in our tests.
Recall that data set twelve’s original and secondary
subjective data both used the SSCQE method. Figure 5 4.3 Predicting order #2 from order #1 plus memory
plots the correlation (y-axis) between the original and When considering Figure 5, the correlations were based
secondary subjective ratings for clips in data set twelve, on just twenty video clips (i.e., five 9-second scenes and
using just a single SSCQE sample taken x seconds into the four systems under test). Therefore, these results do not
nine second clips (x-axis). For example, the correlation constitute sufficient proof as to the duration of memory
for x = 1 second corresponds to original and secondary effects. Let us now consider a different analysis that
viewers having viewed only one second of common video utilizes many more data points. Recall that our secondary
for each of the 20 sequences in data set twelve. test was designed with two randomized viewing orders
(order #1 and order #2), each of 30 minutes duration. The
approach is visually depicted in Figure 6 and described as
follows:
(1) For each second of video in order #1, find the
corresponding second of video in order #2 that contains
that same video content.
(2) Looking at the MOS time histories, build a linear
regression to predict the current MOS of order #2 (value
D) using the matching sample of order #1 (value A), the
delta between matching sample A and the rating ‘Time
Ago’ seconds prior (the length of B), and the delta
between order #1 and #2 ratings at ‘Time Ago’ seconds
prior (the length of C). Note that if ‘Time Ago’ = 0, the
prediction should be perfect.

Figure 5. Correlation of original & secondary ratings using a A
single SSCQE sample extracted from each clip in data
B
set twelve.
Order #1
The minimum correlation level is influenced by the design
of the original and secondary tests. Recall that for both C
the original and secondary experiments, the 9 second
sequences were part of a longer sequence (i.e., a 1 minute D
clip for the original, and a 45-second super-scene for the Order #2
secondary). Since these longer sequences were shown to Time Ago
viewers together for each system under test, the rating of
the content prior to the beginning of the 9 second Figure 6. Predicting order #2: D = w0 + w1A + w2B + w3C.
sequences was (for three out of five scenes) created using
the same video system. The two scenes for which the (3) Use the resulting linear prediction model to assess how
previous content was uncorrelated were (1) the first clip much of order #2’s subjective rating is determined by
displayed in the secondary subjective test’s super-scene; order #1’s subjective rating and the influence of past
and (2) a 9 second clip that was taken from the very events. Value A indicates the ability of the order #1 MOS
beginning of one of data set twelve’s 1 minute video to predict the order #2 MOS. Value B is calculated within
sequences. For the other three scenes, the prior content order #1 as the current value minus the value at Time Ago
was always from the same video system under test. seconds. Value B’s role is to eliminate order #1’s
memory effect from the model. Value C is calculated at
Between 1 and 3 seconds, the correlation is low. Between
Time Ago seconds as order #2 minus order #1. Value C’s
7 and 9 seconds, the original and secondary data
role is to introduce order #2’s memory effect into the
correlations are stable and relatively high (between 0.924
model.
and 0.944). This plot indicates that on average, viewers

Linear prediction weights (w1, w2, and w3) for the model twelve, comparing its order #1 to its order #2 (this original
described in Figure 6 are plotted in Figure 7 as a function experiment was conducted in a similar manner to the
of ‘Time Ago’. The constant term (w0) has been left out secondary experiment, using two randomized viewing
for simplicity. For Figure 7, the current samples (A and orders). The fourth model, labeled ‘Original, dancers’, is
D) are limited to the last sample in each of the 185 video trained over the original subjective data set twelve’s scene
clips, excluding the reference video sequences. Any term ‘dancers’. This video sequence unambiguously contains 1
for which a 95% hypothesis test concludes that the term is minute of continuous content. Notice that, for all four
statistically equivalent to zero has been plotted as being models, the correlation drops to a minimum level after
exactly zero. The w1 term varies around 1.0, indicating approximately 15 seconds, which gives further evidence
that order #2’s subjective data is on average quite similar that video presented earlier than 15 seconds ago has little
to order #1’s subjective data. The w2 term is generally impact on the current SSCQE score.
negative, which has the effect of removing order #1’s
memory effect. The w3 term is generally positive, which
has the effect of introducing order #2’s memory effect.
The w2 and w3 terms drop out of the equation after a
‘Time Ago’ of about 14 seconds, leaving only the w0 and
w1 terms. The w2 term appears again around 90 seconds
later, due to the design of the secondary test which was
based on 45 second long super-scenes (i.e., while the same
super-scene is never presented twice in a row, it can be
presented again after some other super-scene is
presented). From this analysis, the viewer’s current score
does not appear to be influenced by video seen more than
14 seconds ago.

Figure 8. Influence of past events on current viewer opinion
scores.

4.4 Summary
The analyses indicate that the viewers who participated in
our SSCQE experiments considered at most the last 9 to
15 seconds of video when forming their quality estimate.
This is not to say that long sequences are completely
without other merits. Nonetheless, when long video
sequences are used in SSCQE tests, test designers should
not necessarily expect a panel of viewers to rate the video
inherently differently than if shorter sequences are used.
Figure 7. Linear predictor weights as a function of ‘Time Ago’.
5. CONCLUSIONS
Figure 8 shows the correlation between the actual and
SSCQE with hidden reference removal and multiple
predicted order #2 viewer opinion scores when weights
randomized viewer orderings (at least two) can produce
are selected in four different ways. Model ‘Secondary,
quality estimates comparable to DSCQS and DSCS. The
end’ is trained over the secondary subjective data, limited
best method for performing this translation is to take the
to predicting the score at the end of each impaired video
last SSCQE time sample of the scene and compare that
clip. This model uses 185 data points and corresponds to
with the DSCQS or DSCS value. This reassuring result
Figure 7 above. Model ‘Secondary, all’ is trained over
shows that viewers perform essentially the same error
the entire half hour of continuous secondary subjective
pooling function (i.e., the judgment process where
data. The third model, labeled ‘Original, data set 12’, is
perceived errors distributed in space and time are mapped
trained over the entirety of the original subjective data set

to overall estimates of perceived quality) in SSCQE, [7] J. Neter, M. Kutner, C. Nachtsheim, and W.
DSCQS, and DSCS tests. Wasserman, Applied Linear Statistical Models,
Fourth Edition, New York: The McGraw-Hill
While the amount of prior video that is used to form the
Companies, Inc, 1996.
SSCQE quality estimate might be dependent on training
instructions, our SSCQE test subjects appeared to utilize
at most the previous 9 to 15 seconds of video. This fact,
together with the demonstrated ability of SSCQE with
hidden reference removal to replicate double stimulus
testing results, has ramifications for simplifying subjective
test design. Properly designed SSCQE testing (with short
9 to 15 second sequences) may be an effective substitute
for more complicated DSCQS testing. The advantages of
using SSCQE as a substitute would include faster testing
(or more clips rated for the same amount of viewing time
spent) and less viewer fatigue.

6. ACKNOWLEDGMENTS
The authors gratefully acknowledge the contributions of
Philip J. Corriveau for his suggestions regarding SSCQE
with hidden reference removal and comparisons of this
form of testing to DSCQS, Stephen Voran for designing
SSCQE testing devices for our subjective testing
laboratory, and Paul Lemmon for constructing SSCQE
testing devices and conducting the secondary subjective
experiment.

7. REFERENCES
[1] ITU-R Recommendation BT.500-11, “Methodology
for the subjective assessment of the quality of
television pictures,” Geneva, 2002 (available at
www.itu.org).
[2] Video Quality Experts Group (VQEG), “RRNR-TV
Test Plan” (available at www.vqeg.org).
[3] ITU-T COM 9-80-E, “Final report from the video
quality experts group (VQEG) on the validation of
objective models of video quality assessment,”
approved Mar. 2000, Ottawa, Canada (available at
www.vqeg.org).
[4] S. Wolf and M. Pinson, “Video Quality Measurement
Techniques,” NTIA Report 02-392, Jun. 2002
(available at
www.its.bldrdoc.gov/n3/video/documents.htm).
[5] R. Hamberg and H. Ridder, “Time-varying Image
Quality: Modeling the Relation between
Instantaneous and Overall Quality,” SMPTE Journal,
Nov. 1999, p. 802-811.
[6] R. Aldridge, D. Hands, D. Pearson and N. Lodge,
“Continuous quality assessment of digitally-coded
television pictures,” IEEE Proceedings online no.
19981843, Oct. 1997.

Video quality testing methods

More Related Content

Similar to Video quality testing methods (14)

Video quality testing methods