Video Quality Analysis for Concert Video Mashup Generation - Advanced Concepts for Intelligent Vision Systems - page 9

(d)

0.5

(e)

0.4

(c)
(b)

0.3

0.2

(f)

0.1

(a)

0

5740 5760 5780 5800 5820 5840 5860 5880 5900

Frame number

Fig. 5. Quality score of recordings from a three-camera recording, given by different
colors, in a common time-line represented by the frame numbers. Frames indicated by
' ∗ 'areshowninFig.6.

followed by a questionnaire. We ask the participants to indicate the level of the
perceived image quality and camera stability in a 7 point Likert scale, used ex-
tensively in perception tests. The scale is comparable to the mean opinion score
employed in subjective analysis of the television pictures [11]. The measures
image quality and camera stability are chosen from a pilot user-study, as indi-
vidual quality factors like blurriness and blockiness are dicult to differentiate
and evaluate for a general user.
The mean scores of the mashups generated by random, manual and our
quality-based methods in terms of perceived camera stability and image quality
in Fig. 7(a) and (b), respectively. The confidence intervals are presented graph-
ically as an error bar on the mean score, such that if the test is repeated with
other participants from the same target group, there is 95% probability that the
mean score will remain within the interval. The scores are further analyzed to
verify whether the differences between the mean scores are statistically signifi-
cant. We apply a two-way analysis of variance (ANOVA) with repeated measures.
The method (random, quality-based, manual) and concert (C1-C3) are treated
as within-subject independent variables and the response of the participants are
treated as a dependent variable. The results are presented in terms of F-statistic
and p-value such that if p< 0 . 05 there is 95% confidence that the means are
significantly different. Since ANOVA indicates if the means are significantly dif-
ferent, but it does not distinguish which means are different, an additional Tukey
post-hoc test is performed on both independent variables. The results provide
pairwise comparisons of the means and their confidence intervals.
As shown in Fig. 7(a), camera stability score of the manual mashups appear
to be higher than the mashups generated by other methods in all three concerts.
According to the ANOVA analysis, a significant main effect is found for methods
( F =4 . 593 ,p =0 . 010) and for concerts ( F =31 . 853 ,p< 0 . 001). A Tukey test
on method shows that score of the random mashup is significantly lower than the
mashups generated by other methods. Similarly, a Tukey test on concert shows
that C1 is significantly different from C2 and C3.