Video Transcoding in H.264

1. PISATEL Video Transcoding in H.264 Prof. Maurizio Bonuccelli Francesca Martelli Francesca Lonetti

2. PISATEL Outline • H.264 coding and transcoding • Motion Vector Composition algorithms • Overview of our temporal transcoding results (MPEG4, H.263)

3. PISATEL H.264 coding H.264 mean features: • Variable and small block sizes (4×4) • Quarter-pixel resolution • NAL units • New entropic coding algorithms (CABAC) • Deblocking filter • Moto compensation with multiple reference pictures Average bit rate reduction! High complexity!

4. H.264 temporal transcoding Skipping frames to reduce the output bit-rate Three main issues: Motion Vector Computation(MVC) Prediction Error Computation Frame Skipping Policy 3 2 1 PISATEL 3 1 3 2 1 Transcoder Output bit-rate 32 Kb Input bit-rate 64 Kb

5. MVC F(n-2) F(n-1) F(n) mvn mvn-1 skipped Motion Vector Composition • The motion vectors are computed by • Motion Vector Composition Algorithms (BI, TVC, FDVS, ADVS) • Restricted Motion Estimation (RME)

6. MVTVC=MV1n-1 Telescopic Vector Composition MVn (skipped) MV1n-1 MV2n-1 MB1n-1 MBn MB2n-1 MV4n-1 MBn MV3n-1 MB3n-1 MB4n-1 F(n) F(n-1) New reference area MVn=MVn+MVTVC F(n-2)

7. Bilinear Interpolation MVBI=(1-α)(1-β)MV1n-1+(α)(1-β)MV2n-1+(1-α)(β)MV3n-1+(α)(β)MV4n-1 MVn (skipped) MBn MV2n-1 MB1n-1 α MB2n-1 MV1n-1 β MV4n-1 MVBI MBn MV3n-1 F(n) MB4n-1 MB3n-1 New reference area F(n-1) MVn=MVn+MVBI F(n-2)

8. Previous standards ? H.264 features • New in H.264: variable macroblock partition (16 motion vectors for each macroblock) • Transcoder keeps the same partitions of the remote encoder (most efficient solution) • How to apply MVC in H.264? • BI and TVC adaptation • New MVC algorithm H.264 ?

9. MVC in H.264 MB1 MB2 Mv_MVC= f(mv_MB1,mv_B2, mv_B3, mv_B4) LIV.0 MB3 MB4 MB1 MB3 MB4 MB2 Mv_MB1= mv_B2 Mv_MB3= f(mv_B2,mv_B4) Mv_MB2= f(mv_B1,mv_B2) Mv_MB4= f(mv_B1,mv_B2) B1 B2 B2 B1 B1 LIV.1 B1 B2 B3 B4 B3 B4 B2 B2 B4 B2 B1 B1 B2 B2 LIV.2 sb1 sb2 LIV.3 Mv_B2 Mv_B2 Mv_B4 Mv_B1 Mv_B2 Mv_B1= f(sb1,sb2) Mv_B2

10. A new MVC algorithm • Basic Idea: looking for a reference area, the most similar to the current macroblock • Two steps: • For each motion vector of a macroblock (or block) of the reference area in the skipped frame,compute the difference between the macroblock and an area pointed by the motion vector with the same size of the current macroblock • Choose the motion vector of the reference area that minimizes this difference

11. New MVC algorithm: example F(n-2) F(n-1) F(n) mv1 mv2 MB1 MB2 A1 mv3 b1 b2 MB A2 mv4 mv5 b4 b3 N A A3 mv A5 b1 mv7 M b1 b2 b2 mv8 A4 A6 new_mv A8 A7 MB3 MB4 mv6 (skipped) Vf=argmin i Є SMSE(A, Ai) = argmin i Є S(1/NxM |A-Ai|2)

12. MVC performance (time) TVC BI New ME PISATEL

13. Motion Estimation New MVC Algorithm MVC performance (quality) Akiyo PSNR(dB) Frames

14. New MVC Algorithm BI Algorithm TVC Algorithm MVC performance (quality) Akiyo PSNR(dB) Frames

15. Motion Estimation New MVC Algorithm MVC performance (quality) Coastguard PSNR(dB) Frames

16. New MVC Algorithm BI Algorithm TVC Algorithm MVC performance (quality) Coastguard PSNR(dB) Frames

17. Our activities in video transcoding • We started in studying video features in MPEG4 • With this codec, we developed two temporal transcoders: • One based on an architecture known as FSC (Frame Skipping Control) • One based on an architecture known as DFS (Dynamic Frame Skipping) • By simulation, we realized that the second architecture results better in terms of video quality, while the first one results better in terms of processing times

18. MPEG4 transcoder architectures • After a skipped frame: • In the FSC architecture, motion vectors and prediction errors are computed without executing motion estimation  worse quality, re-encoding errors amassment • In the DFS architecture, motion vectors and prediction errors are computed with a restricted motion estimation  better quality, but worse time performance

19. Motion Vector Composition • In literature we found 4 MVC algorithms: • Bilinear interpolation • Telescopic Vector Composition • Forward Dominant Vector Selection • Activity Dominant Vector Selection • By using the DFS architecture (that performs motion estimation) we realized by simulation that these algorithms are equivalent

20. Architectures and MVC algorithms M. A. Bonuccelli, F. Lonetti, F. Martelli. Video Transcoding Architectures for Multimedia Real Time Services, ERCIM News No. 62, pp. 39-40, July 2005.

21. Skipping policies • We developed 4 skipping policies, all of them based on the transcoder buffer occupancy: • Buffer based • Motion activity • Consecutive skipping • Random • M. A. Bonuccelli, F. Lonetti, F. Martelli. Temporal Transcoding for Mobile Video Communication. In Proceedings of 2nd Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services (Mobiquitous 2005), pp.502-506, July 17-21, 2005, San Diego, CA.

22. H.263 Video Transcoder • Then, we studied the H.263 video codec, and we developed a temporal transcoder, based on DFS architecture. • By observing the test results, we realize that video quality is influenced by the rate control of the front encoder • Then, we tested and implemented several rate control algorithms for the front encoder: • TMN5 • TMN8 •  domain • Perceptual rate control • Multiple zone (Activity)

23. H.263 Video Transcoder • In our transcoder architecture, every frame is first transcoded, and then transmitted or dropped. • To avoid the computation of frames that will be discarded, we developed a new frame skipping policy that predicts the frames to be transcoded • M. A. Bonuccelli, F. Lonetti, F. Martelli. A Fast Skipping Policy for H.263 Video Transcoder. In Proceedings of 12th International Workshop on Systems, Signals and Image Processing (IWSSIP'05). September 22-24, 2005, Chalkida, Greece.

24. Finally, H.264 • This codec is the product of the union between VCEG (H.263) and MPEG, forming the Joint VideoTeam (JVT) • It includes all benefits of previous standards in order to achieve good quality performance even at low bit-rate • We worked in two directions: • Optimizing the codec, to speed up it • Implementing a transcoder

25. Optimizing H.264 encoder • We operated some modifications to the reference software in order to obtain acceptable encoding times: • instead of computing all half and quarter pixels in two rounds, we compute them in only one round • fast way for choosing the optimal partitioning: instead of using the SAD (Sum of Absolute Differences) measure as decision parameter, we use other metrics: • the number of differences in terms of pixels • the maximum difference value • the average difference value • the most popular difference value compared with proper self-adjusting thresholds.

26. H.264 transcoder • We first implemented the cascade pixel domain transcoder by simply concatenating a decoder with an encoder, in order to develop the motion vector composition algorithms • Then we are implementing the frame skipping to obtain a transcoder that decides the frames to be dropped

27. H.264 rate control • Finally, we are implementing the TMN8 rate control algorithm to be used in the front encoder • We think that, with a rate control algorithm able to skip frame in encoding phase, the transcoding process may be improved both in terms of quality and computation time

28. Master Theses • Luigi D’Amaro. Algoritmi per la transcodifica video. • Gianni Rosa. Transcodifica video per comunicazione mobile: studio di rate control. • Luca Leonardi. Transcodifica video temporale: politiche di frame skipping. • Marina Paletta. Realizzazione di un transcodificatore video temporale H.264 per video comunicazione mobile. • Riccardo Vagli. Implementazione di un transcoder video basato sullo standard H.264/AVC. • Alsona Dema. Rate Control in H.264.

29. Conclusions • We studied the video transcoding problem in real-time communications • We developed temporal transcoders with MPEG4, H.263 and H.264 codecs • We developed some skipping policies to be used in each transcoder • We developed three MVC algorithms to be used in the H.264 transcoder

30. Acknowledgements • We thank all ERI people who introduced us in this research area, for the helpful discussions and advices • We thank PisaTel Lab people • We thank all students who worked with us in this project