Video Codecs Comparison Part 3: -accurate Comparison Project head: Dmitriy Vatolin Testing: Sergey Grishin Translating: Daria Kalinkina, Stanislav Soldatov Preparing: Nikolai Trunichkin 9 testing sequences! 11 days (26 hours) total compression time! 33 tested codecs! 243 resulting sequences! May 23 CS MSU Graphics&Media Lab Video Group http://www.compression.ru/video/
VIDEO GROUP MOSCOW, 15 MAY 23 Video codecs comparison Part 3: -accurate Comparison 15 May 23 Contents Contents... 2 Divx 4.2 and Divx 3.1 fast motion (battle 134kbps)... 3 Divx 4.2 and Divx 5.2 (battle 46кбит/с)... 6 Divx 4.2 and Microsoft 3688 v3 (battle 134 kbps)... 9 Divx 4.2 and VP 3.1 (bankomatdi 34 kbps)... 12 Divx 4.2 and Xvid 2.1 (battle 184 kbps)... 14 Divx 4.2 and MorganMultimedia JPEG2 (battle 134 kbps)... 17 Divx 4.2 and VSS H.264 (battle 134 kbps)... 2 Outline... 23 CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 2
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and Divx 3.1 fast motion (battle 134kbps) 2.5. 1 4 's size 2. 1 4 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 2 4 6 8 1 12 14 16 Divx 3.1 fast motion Divx 4.2 5 's psnr 45 Y-PSNR 4 aver_psnr 35 3 2 4 6 8 1 12 14 16 Divx 3.1 fast motion Divx 4.2 1.8. 1 4 By size difference, Bytes Size difference 1.8. 1 4 2 4 6 8 1 12 14 16 Size(Divx 3.1 fm) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 3
VIDEO GROUP MOSCOW, 15 MAY 23 1.5. 1 4 's size 1. 1 4 aver_size1 aver_size 5 9 95 1 15 11 115 Divx 3.1 fast motion Divx 4.2 's psnr 45 Y-PSNR 4 aver_psnr 35 8 85 9 95 1 15 11 115 Divx 3.1 fast motion Divx 4.2 Y-YUV PSNR difference By Y-YUV psnr difference, db 4 4 2 4 6 8 1 12 14 16 Y(Divx 3.1) - Y(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 4
VIDEO GROUP MOSCOW, 15 MAY 23 U-YUV PSNR difference By U-YUV psnr difference, db 4 4 2 4 6 8 1 12 14 16 U(Divx 3.1 fm) - U(Divx 4.2) Conclusions: Divx 3.1 has a higher metric than Divx 4.2; this can be clearly seen on the digrams 2 and 5. Both codecs keep the U- and V-components almost similarly (see the last diagram). CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 5
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and Divx 5.2 (battle 46кбит/с) 's size 1. 1 4 8 6 4 aver_size1 aver_size 2 2 4 6 8 1 12 14 16 Divx 5.2 Divx 4.2 45 's psnr 4 Y-PSNR aver_psnr 35 3 2 4 6 8 1 12 14 16 Divx 5.2 Divx 4.2 1315 By size difference, Bytes Size difference 1315 2 4 6 8 1 12 14 16 Size(Divx 5.2) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 6
VIDEO GROUP MOSCOW, 15 MAY 23 's size 1. 1 4 8 6 4 aver_size1 aver_size 2 8 82 84 86 88 9 92 94 96 98 1 Divx 5.2 Divx 4.2 's psnr 4 Y-PSNR 35 aver_psnr 8 82 84 86 88 9 92 94 96 98 1 Divx 5.2 Divx 4.2 Y-YUV PSNR difference By Y-YUV psnr difference, db 1 1 2 4 6 8 1 12 14 16 Y(Divx 5.2) - Y(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 7
VIDEO GROUP MOSCOW, 15 MAY 23 U-YUV PSNR difference 1 By U-YUV psnr difference, db 1 2 4 6 8 1 12 14 16 U(Divx 5.2) - U(Divx 4.2) According to the first diagram Divx 5.2 has a better Y-PSNR metric; according to the second one Divx 4.2 has a better U-PSNR metric. That means that the developers of the Divx 5.2 codec improved the Y-component at the expense of the U- and V-components, which seems to be the right decision because the human eye is more sensitive to changes in brightness than to changes in color. Also it should be mentioned that these codecs have approximately the same frame size. It can be especially well seen on diagram 5 (which is an increased diagram 4). The dashed lines mean the average frame size values for each of the two codecs; here these lines flow together into one line. The last diagram represents the difference in Y-PSNR metric. CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 8
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and Microsoft 3688 v3 (battle 134 kbps) 3. 1 4 's size 2.5. 1 4 2.1 4 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 2 4 6 8 1 12 14 16 Microsoft v3 Divx 4.2 55 's psnr 5 Y-PSNR 45 4 aver_psnr 35 3 2 4 6 8 1 12 14 16 Microsoft v3 Divx 4.2 1.31.1 4 By size difference, Bytes Size difference 1.31. 1 4 2 4 6 8 1 12 14 16 Size(Microsoft v3) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 9
VIDEO GROUP MOSCOW, 15 MAY 23 2. 1 4 's size 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 8 82 84 86 88 9 92 94 96 98 1 Microsoft v3 Divx 4.2 's psnr 5 Y-PSNR 45 4 aver_psnr 35 8 82 84 86 88 9 92 94 96 98 1 Microsoft v3 Divx 4.2 Y-YUV PSNR difference By Y-YUV psnr difference, db 8 8 2 4 6 8 1 12 14 16 Y(Micorosoft v3) - Y(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 1
VIDEO GROUP MOSCOW, 15 MAY 23 U-YUV PSNR difference By U-YUV psnr difference, db 4 4 2 4 6 8 1 12 14 16 U(Microsoft v3) - U(Divx 4.2) Conclusions: Microsoft v3 has a noticeably higher average metric value than Divx 4.2. Microsoft v3 better keeps the U-component. Divx 4.2 better keeps the Y-component; its Y-PSNR diagram has smaller amplitude. The second diagram represents the sharp increase of the Microsoft v3 codec s Y-PSNR near the 8th frame. This increase can be seen by sight during the playback. CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 11
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and VP 3.1 (bankomatdi 34 kbps) 's size 1. 1 4 8 6 4 aver_size1 aver_size 2 5 1 15 2 25 3 35 4 VP 3.1 Divx 4.2 42 's psnr 4 38 Y-PSNR 36 34 32 aver_psnr 3 28 5 1 15 2 25 3 35 4 VP 3.1 Divx 4.2 3.71.1 4 By size difference, Bytes Size difference 3.71. 1 4 5 1 15 2 25 3 35 4 Size(VP 3.1) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 12
VIDEO GROUP MOSCOW, 15 MAY 23 1. 1 4 's size 8 6 4 aver_size1 aver_size 2 2 4 6 8 1 12 14 16 VP 3.1 Divx 4.2 Y-YUV PSNR difference By Y-YUV psnr difference, db 8 8 5 1 15 2 25 3 35 4 Y(VP 3.1) - Y(Divx 4.2) U-YUV PSNR difference 7 By U-YUV psnr difference, db 7 5 1 15 2 25 3 35 4 U(VP 3.1) - U(Divx 4.2) Conclusions: Div 4.2 better keeps the Y-component. Both codecs keep the U-components similarly. Peaks on the diagram for VP 3.1 correspond with the key-frames. So keyframes are the only frames, where PSNR of VP 3.1 is higher than PSNR of Divx 4.2. Changes of metric with such amplitude can be seen by sight. It means that one can see changes in image quality during the playback. Also key-frames usage causes a significant increase of the average frame size, so here it s not justified. Without generating key-frames VP 3.1 would gain much benefit on bitrate. Periodical rises on the diagram are caused not by the codec s but by the clip s specificity. The bankomat clip contains periodically appearing sequences. CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 13
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and Xvid 2.1 (battle 184 kbps) 4. 1 4 's size 3. 1 4 2. 1 4 aver_size1 aver_size 1. 1 4 2 4 6 8 1 12 14 16 Xvid Divx 4.2 44 's psnr 42 Y-PSNR 4 38 aver_psnr 36 34 32 2 4 6 8 1 12 14 16 Xvid Divx 4.2 1.35. 1 4 By size difference, Bytes Size difference 1.35. 1 4 2 4 6 8 1 12 14 16 Size(Xvid) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 14
VIDEO GROUP MOSCOW, 15 MAY 23 's size 2.5. 1 4 2. 1 4 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 8 82 84 86 88 9 92 94 96 98 1 Xvid Divx 4.2 's psnr 4 aver_psnr Y-PSNR 35 8 82 84 86 88 9 92 94 96 98 1 Xvid Divx 4.2 Y-YUV PSNR difference By Y-YUV psnr difference, db 4 4 2 4 6 8 1 12 14 16 Y(Xvid) - Y(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 15
VIDEO GROUP MOSCOW, 15 MAY 23 U-YUV PSNR difference By U-YUV psnr difference, db 8 8 2 4 6 8 1 12 14 16 U(Xvid) - U(Divx 4.2) Conclusions: Y-PSNR of the frame after Divx 4.2 is 2 db worse. Divx 4.2 better keeps the U-component. Xvid better keeps the metric and the frame size; oscillations on the corresponding diagrams have smaller amplitude. CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 16
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and MorganMultimedia JPEG2 (battle 134 kbps) 3. 1 4 's size 2.5. 1 4 2.1 4 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 2 4 6 8 1 12 14 16 JPEG2 Divx 4.2 55 's psnr 5 45 Y-PSNR 4 aver_psnr 35 3 25 2 4 6 8 1 12 14 16 JPEG2 Divx 4.2 1.66.1 4 By size difference, Bytes Size difference 1.66. 1 4 2 4 6 8 1 12 14 16 Size(JPEG2) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 17
VIDEO GROUP MOSCOW, 15 MAY 23 2.1 4 's size 1.5.1 4 1. 1 4 aver_size1 aver_size 5 8 82 84 86 88 9 92 94 96 98 1 JPEG2 Divx 4.2 's psnr 4 Y-PSNR aver_psnr 35 8 82 84 86 88 9 92 94 96 98 1 JPEG2 Divx 4.2 Y-YUV PSNR difference By Y-YUV psnr difference, db 12 12 2 4 6 8 1 12 14 16 Y(JPEG2) - Y(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 18
VIDEO GROUP MOSCOW, 15 MAY 23 U-YUV PSNR difference By U-YUV psnr difference, db 8 8 2 4 6 8 1 12 14 16 U(JPEG2) - U(Divx 4.2) Conclusions: JPEG2 perfectly keeps the bitrate (see picture 1). JPEG2 better keeps the metric on the dynamic scenes, although it is significantly inferior to Divx 4.2 on the static ones. JPEG2 better keeps the U-component. The difference in their average metric indexes is approximately.5 db. That means that being superior in the frame size JPEG2 also has a bit better quality. Despite the constant frame size, the metric index of JPEG2 changes greatly. According to the second diagram the amplitude of its oscillations reaches 1 db. These metric overfalls evidently badly affect the quality of the compressed sequence. Divx 4.2 keeps the metric much more stably. CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 19
VIDEO GROUP MOSCOW, 15 MAY 23 Divx 4.2 and VSS H.264 (battle 134 kbps) 3. 1 4 's size 2.5. 1 4 2.1 4 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 2 4 6 8 1 12 14 16 VSS H.264 Divx 4.2 55 's psnr 5 PSNR 45 4 aver_psnr 35 3 2 4 6 8 1 12 14 16 VSS H.264 Divx 4.2 2.21.1 4 By size difference, B Size delta 2.21. 1 4 2 4 6 8 1 12 14 16 Size(VSS H.264) - Size(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 2
VIDEO GROUP MOSCOW, 15 MAY 23 3. 1 4 's size 2.5. 1 4 2.1 4 1.5. 1 4 1. 1 4 aver_size1 aver_size 5 8 82 84 86 88 9 92 94 96 98 1 VSS H.264 Divx 4.2 55 's psnr 5 PSNR 45 4 aver_psnr 35 3 8 82 84 86 88 9 92 94 96 98 1 VSS H.264 Divx 4.2 1 By Y-YUV psnr difference, db Y-YUV PSNR delta 1 2 4 6 8 1 12 14 16 Y(VSS H.264) - Y(Divx 4.2) CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 21
VIDEO GROUP MOSCOW, 15 MAY 23 5 By U-YUV psnr difference, db U-YUV PSNR delta 5 2 4 6 8 1 12 14 16 U(VSS H.264) - U(Divx 4.2) Conclusions: Y-PSNR of H.264 is on average 1 db higher than the one of Divx 4.2. Four I-frames, which are characterized by the sudden changes in Y-PSNR, can be seen on the Y-PSNR diagrams for VSS H.264 (1-16 frames interval). These overfalls can be seen by sight in the video sequence. It should be mentioned that the overalls on the corresponding frames after Divx 4.2 are 1 db less (see the next to last picture). Divx 4.2 keeps the U-component a little better. CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 22
VIDEO GROUP MOSCOW, 15 MAY 23 Outline Video Codecs Comparison consists of the following sections: Part 1: Methodology Part 2: PSNR Diagrams For All Video Codecs Part 3: -accurate Comparison this document Part 4: Visual Comparison NOTE: These files contain only a VERY SMALL PART of the processed and measured data. If you find an error in this document, please write to video@graphics.cs.msu.su For new materials please check http://compression.ru/video/ CS MSU G&M Lab/ Video Group / HTTP://WWW.COMPRESSION.RU/VIDEO/ 23