Page 115 - ITU Journal, ICT Discoveries, Volume 3, No. 1, June 2020 Special issue: The future of video and immersive media

ITU Journal: ICT Discoveries, Vol. 3(1), June 2020

s
44 44
PSNR-YUV Min PSNR-YUV Min
H(0,0)=2 H(0,1)=3 H(0,2)=2 H(0,3)=3 H(0,4)=2 42 PSNR-YUV Max 42 PSNR-YUV Max
PSNR-YUV
PSNR-YUV
40 PSNR-YUV + SD 40 38 PSNR-YUV + SD
PSNR-YUV - SD
PSNR-YUV - SD
PSNR-YUV (dB) 36 PSNR-YUV (dB) 36 34
38
34
H(1,0)=3 H(1,1)=3 H(1,2)=3 H(1,3)=3 H(1,4)=3 32 32
30
28 30 28
26 26
H(2,0)=2 H(2,1)=3 H(2,2)=1 H(2,3)=3 H(2,4)=2
24 24
10 -4 10 -3 10 -2 10 -1 10 0 10 1 10 -4 10 -3 10 -2 10 -1 10 0 10 1
bit rate (bpp) bit rate (bpp)
(a) Bikes (b) Danger
H(3,0)=3 H(3,1)=3 H(3,2)=3 H(3,3)=3 H(3,4)=3
44 46
PSNR-YUV Min PSNR-YUV Min
42 PSNR-YUV Max 44 PSNR-YUV Max
PSNR-YUV PSNR-YUV
H(4,0)=2 H(4,1)=3 H(4,2)=2 H(4,3)=3 H(4,4)=2 40 PSNR-YUV + SD 42 40 PSNR-YUV + SD
PSNR-YUV - SD
PSNR-YUV - SD
PSNR-YUV (dB) 36 PSNR-YUV (dB) 38 36
38
34
Reference views 32 34 32
30
t 28 30
Intermediate views 26 28
24 26
10 -4 10 -3 10 -2 10 -1 10 0 10 1 10 -4 10 -3 10 -2 10 -1 10 0 10 1
bit rate (bpp) bit rate (bpp)
Fig. 8 – 4DPM hierarchical partitioning into reference and intermediate
views, with ℎ = 3, when processing the views ( , ) of for which (c) Fountain (d) Pillars
( , ) = 3.
Fig. 9 – 4DTM: PSNR-YUV max, min, average and standard deviation
JPEG Pleno database [18] to fairly evaluate experimental values of the reconstructed views.
results of the different proposed codecs. All the experi-
mental comparisons performed in this paper followed the 52 VM 2.0 (4DPM) 44 VM 2.0 (4DPM)
50 x265 42 x265
procedures as outlined in the JPEG Pleno Common Test 48 40
Conditions (CTC) document [14], which also defines all 46 44 38 36
the metrics used for the evaluation (PSNR-Y, PSNR-YUV PSNR-YUV (dB) 42 PSNR-YUV (dB) 34
and SSIM). For completeness, we specify below the equa- 40 38 32 30
tion that was used to calculate the PSNR-YUV, this formula 36 28
is specified in [14] and [19]: 34 26
32 24
10 -3 10 -2 10 -1 10 0 10 -3 10 -2 10 -1 10 0
= (6 × + + )/8 bit rate (bpp) bit rate (bpp)

(a) Greek (b) Sideboard
Also, the precise configuration of the x265 codec is de-
fined in the CTC document. Please note that x265 was 48 VM 2.0 (4DPM) 46 VM 2.0 (4DPM)
46 x265 44 x265
used instead of HM reference software since the latter 44 42
was to slow for the experiments to be completed. For the 42 40 40
4DTM codec only results for the lenslets-based data sets PSNR-YUV (dB) 38 PSNR-YUV (dB) 38 36
are reported, as its coding design is suitable for very high- 36 34 34
density light fields. 32 32
30 30
Low variation of the quality of the reconstructed views is 28 10 -3 10 -2 10 -1 10 0 28 10 -3 10 -2 10 -1
bit rate (bpp) bit rate (bpp)
required in refocusing applications, for example. Fig. 9
(c) Tarot (d) Set2
depicts the values of average, maximum, minimum,
and confidence interval of one standard deviation of Fig. 10 – 4DPM: PSNR-YUV values of the reconstructed high density
PSNR-YUV for the data sets Bikes, Danger de Mort, camera-array images.
Fountain&Vincent2 and Stone Pillars Outside, respec-
the lenslets data sets: Bikes, Danger de Mort, Foun-
tively [18]. Due to the lenslet light field camera acqui-
tain&Vincent2 and Stone Pillars Outside [14]. Both cod-
sition process, in all lenslet data sets, the four views at
ing modes outperform the x265 codec implementation
the corners are much darker than the rest. When coding
of the H.265/HEVC standard [14] (Table 1). For these
these data sets using the 4DTM the corners views are mul-
very high-density light fields, the 4DTM presents bit-rate
tiplied by 4 before encoding and divided by 4 after decod-
savings for the same PSNR-YUV values when comparing
ing. This procedure decreases the corner views coding er-
with the 4DPM. The former also outperforms the latter
ror by 4, increasing their metric values. Therefore, in or-
when assessing the BD-PSNR(dB) values for the same av-
der to show fairer results, these corner views have been
erage PSNR-YUV difference in dB for the same bit rate,
removed from the PSNR-YUV Max computation (Fig. 9).
as reported in Table 1. When comparing only the lumi-
nance results (PSNR-Y), the 4DTM is outperformed by the
Table 1 shows the Bjøntegaard delta rate (BD-BR (%))
4DPM in both BD-rate and BD-PSNR values for the Foun-
and the Bjøntegaard delta PSNR (BD-PSNR (dB) [20] re-
tain&Vincent2 data set (Table 1). The 4DTM codec does
garding the PSNR-Y and PSNR-YUV results obtained for
© International Telecommunication Union, 2020 93