Information Technology Reference
In-Depth Information
Ta b l e 2
Objective evaluation of the four frame doubling methods in test: Frame repetition,
frame averaging and variational TSR without/with GCA. MSE and PSNR scores are given for
the four test sequences
Square
,
Cameraman Pan
,
Building
and
Control Panel
.
Square Cameraman Building Control
Pan
Panel
Frame repetition
MSE
158.4
1887.9
462.5
1369.8
PSNR
26.13
15.37
21.48
16.76
Frame averaging
MSE
82.54
1208.3
287.8
849.9
PSNR
28.96
17.31
23.54
18.84
TSR without GCA
MSE
13.05
39.47
13.82
76.59
PSNR
36.97
32.17
36.73
29.29
TSR with GCA
MSE
8.97
107.9
16.44
96.87
PSNR
38.60
27.80
35.97
28.27
means larger panning/tilting motions from frame to frame as we now do 12.5 to 25
fps frame doubling. Since frame repetition is not really worth comparing with other
results in stills, and since its motion portrayal is the same as in the input, we left it
out of the subjective evaluation but have included it here as it is the most widely use
method for frame rate up-conversion.
As the results in Table 2 show, our variational TSR algorithms outperforms frame
averaging as it was also the case in the subjective evaluation. It is also no surprise
that in the presence of motion, frame repetition is clearly the objectively worst per-
forming frame doubling algorithm. Whether it is subjectively worse than frame av-
eraging is however a question up for debate because of the double exposure in new
frame in frame averaging, which introduces additional artifacts.
Returning to the far better variational TSR frame doublers, the use of GCA
helps in the case of
object
motion. Variational TSR with GCA gives the best ob-
jective result on
Square
, which corresponds well with the subjective results. For
the two sequences
Cameraman Pan
and
Building
dominated by global mo-
tions, the non-GCA version is objectively better than the GCA version, which
can be explained by the GCA version tending to overfit the flows. The boundary
problems in
Cameraman Pan
are judged from the objective result worse in the
GCA version. On
Control Panel
the non-GCA version produces a smoother
flow field and thus the intensity output is also somewhat smoother, which helps
dampen the problems with wrong flow estimations of the complex flow in
Control
Panel
.
From our combined tests results, we can conclude that variational TSR with-
out GCA performs slightly better or the same as TSR with GCA in cases where
the sequences are dominated by global flow (camera motion). It is clear that our
motion compensated variational TSR frame doublers are producing outputs far su-
perior to the outputs from the simple methods frame averaging and frame repetition.
