Nine months have passed since my last comparisson of the HEVC, VP9 and Daala compression efficiency. Daala was in the development stage and is still being developed. Recently I've updated the codec to see the changes.
My apologies to the haters of PSNR and the lovers of SSIM. I'm still using PSNR.
The two revisions of Daala codec are compared in the following experiments. The newer one is a master commit on July 15, 2014. The older one is a master commit on October 3, 2014.
First of all intra compression efficiency increased. The following table contains Bjontegaard delta rates and delta PSNRs for the JCT-VC test sequences. I've changed command line interface for the means of automatisation process I'm used to. Daala encoder is configured with daala_info structure. The keyframe_rate parameter is set to 1. The experiments are held for video quality parameters of \(vq=2\cdot n,\) where \(n = [2,\:...,\:16]\).
The changes in key-frame compression efficiency of the newer version comparred to the older one are the following:
Table 1: Daala 15-07-2014 compared to Daala 03-10-2013 on intra-only configuration
Sequence BD-RATE BD-PSNR-Y BD-PSNR-UV
Traffic -23.49 1.71 -0.11
PeopleOnStreet -21.29 1.74 0.22
Kimono -51.30 1.98 0.24
ParkScene -20.44 1.16 0.10
Cactus -18.06 0.90 0.09
BQTerrace -8.31 0.64 0.24
BasketballDrive -32.81 1.15 0.52
RaceHorsesC 2.60 -0.16 -0.05
BQMall -2.87 0.27 0.00
PartyScene 10.75 -0.94 0.09
BasketballDrill -8.00 0.52 0.36
RaceHorses -8.39 -0.88 -0.01
BQSquare 1.48 -18.2 -19.45
BlowingBubbles -7.36 0.59 0.30
BasketballPass -13.00 1.09 0.08
Vidyo1 -28.04 1.98 0.89
Vidyo3 -25.47 1.85 1.51
Vidyo4 -28.21 1.77 1.34
BasketballDrillText -6.19 0.44 0.49
ChinaSpeed 23.83 -2.12 0.41
SlideEditing 22.02 -2.80 -0.07
SlideShow -8.60 1.09 0.16
Generally the compression efficiency increase by 20-30% for the majority of test sequences. However, the compression efficiency of ChinaSpeed and SlideEditing test sequences decresed by 20% (in BD-Rate).
For inter-frame coding efficiency the key frame is set to be only the first one. The rest configuration is the same. And the results follow.
Table 2: Daala 15-07-2014 compared to Daala 03-10-2013 on inter frame configuration
Sequence BD-RATE BD-PSNR-Y BD-PSNR-UV
Traffic -65.10 3.47 1.79
PeopleOnStreet -53.88 3.55 2.58
Kimono -83.29 3.57 1.12
ParkScene -41.18 1.71 1.00
Cactus -43.62 1.64 0.98
BQTerrace -20.72 0.93 1.65
BasketballDrive -64.49 2.24 1.09
RaceHorsesC -24.20 1.53 1.88
BQMall -39.31 2.14 2.73
PartyScene -18.15 1.31 2.37
BasketballDrill -43.91 2.15 2.21
RaceHorses -29.16 2.00 2.45
BQSquare -29.08 2.36 3.33
BlowingBubbles -26.17 1.59 2.20
BasketballPass -57.43 3.83 4.16
Vidyo1 -82.65 4.49 1.97
Vidyo3 -75.76 4.01 1.63
Vidyo4 -78.23 3.95 1.77
BasketballDrillText -45.90 2.41 3.07
ChinaSpeed -36.41 2.90 3.23
SlideEditing -76.28 14.97 13.09
SlideShow -69.87 9.95 8.07
There is a significant compression efficiency gain of 20-80%. About 83% bitrate reduction achieved on Kimono and Vidyo1 test sequences.
However, comparred to the JM reference encoder Daala's performance is still rather poor.
Updated Daala's keyframe compression efficiency against JM v.18.4 "Intra HE" configuration:
Table 3: Daala 15-07-2014 intra compared to JM v.18.4 "Intra HE"
Sequence BD-RATE BD-PSNR-Y BD-PSNR-UV
Traffic 17.65 -0.96 -2.13
PeopleOnStreet 20.53 -1.21 -1.22
Kimono -1.60 0.04 -1.03
ParkScene 17.04 -0.77 -1.27
Cactus 30.30 -1.13 -1.17
BQTerrace 35.12 -2.08 -1.16
BasketballDrive 52.44 -1.13 -2.03
RaceHorsesC 40.90 -2.56 -1.59
BQMall 50.21 -2.83 -1.63
PartyScene 49.14 -3.48 -1.45
BasketballDrill 39.64 -1.81 -2.42
RaceHorses 30.13 -2.99 -2.17
BQSquare 55.50 -9.56 -7.00
BlowingBubbles 37.50 -2.18 -1.77
BasketballPass 39.72 -2.24 -3.13
Vidyo1 42.59 -1.91 -1.42
Vidyo3 59.39 -2.74 -0.80
Vidyo4 43.91 -1.79 -1.25
BasketballDrillText 47.43 -2.25 -3.05
ChinaSpeed 91.41 -6.21 -2.84
SlideEditing 121.43 -10.46 -3.95
SlideShow 270.12 -9.48 -10.55
Obviously Daala provides 1.5 times higher bitrate compared to the JM encoder. Moreover, on the SlideEditing and SlideShow test sequences the bitrate is 2-3 times higher. Intra compression efficiency of the Daala coder seems to remain not efficient enough.
Table 4: Daala 15-07-2014 inter compared to JM v.18.4 "RA B HE"
Sequence BD-RATE BD-PSNR-Y BD-PSNR-UV
Traffic 149.34 -3.12 -3.51
PeopleOnStreet 68.13 -2.46 -2.33
Kimono 3.37 -0.13 -2.48
ParkScene 149.09 -3.15 -3.25
Cactus 199.99 -2.43 -2.04
BQTerrace 297.19 -3.23 -2.48
BasketballDrive 87.29 -1.29 -2.78
RaceHorsesC 110.55 -3.60 -2.70
BQMall 200.56 -4.70 -3.10
PartyScene 188.62 -5.70 -3.32
BasketballDrill 173.82 -4.06 -4.43
RaceHorses 113.98 -4.32 -3.73
BQSquare 369.81 -7.43 -2.80
BlowingBubbles 172.57 -4.56 -3.91
BasketballPass 151.03 -4.59 -4.43
Vidyo1 302.75 -3.61 -2.92
Vidyo3 243.38 -3.62 -2.57
Vidyo4 241.21 -3.06 -3.27
BasketballDrillText 193.53 -4.59 -5.34
ChinaSpeed 287.76 -9.32 -5.97
SlideEditing 378.40 -21.32 -11.34
SlideShow 382.95 -14.45 -18.80
The inter-frame compression performance of Daala is compared to the JM
v.18.4 in "Random Access B HE" configuration. The overall compression
efficiency of the Daala video encoder is 2-5 times lower than the
compression efficiency of the JM reference encoder.
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment