近年來遊戲實況串流平台大為流行，然而從近來的研究中顯示這一
些平台需要使用大量的網路流量，間接的導致不易提高使用者體驗。
在這一篇論文中我們提出了使用『感興趣片段』之概念的遊戲實況串
流平台， 並以此概念對平台進行優化。我們的平台使用從實況主以
及觀眾電腦中收集來的特徵， 結合當前最優秀之機器學習模型來自動
化的偵測目前實況影片的片段是不是會吸引觀眾的注意。在決定了
當前實況影片片段的重要性之後， 系統中可以使用的頻寬會以一個利
用『資料率失真理論』進行最佳化的方式來分配給感興趣的觀眾， 其
中會使用影片的片段重要性來作為要傳輸的影片品質依據。在這個系
統運作背後的理念為：唯有在觀眾正在注意實況影片時，才會發生降
低使用者體驗的情況。在我們的使用從現實世界中收集來的數據進行
的模擬顯示，在使用10次交叉驗證的條件下， 我們的SoI偵測演算法
在使用分類器偵測『感興趣片段』的F度量高達0.96，在使用迴歸器下
的平方度量高達0.87。而在針對資源分配的模擬實驗中我們的資源分
配演算法可以達到：(i)增進影片品質高達5 dB， (ii) 最高可以節省50
Gbps的頻寬使用率，以及(iii)該演算法可以有效率的完成資源分配決
定， 並且可以承受巨量的使用者。我們在本篇論文中介紹的平台是一
個開源平台， 可以被眾多研究者以及工程師使用來改進現有的遊戲實
況串流平台。

Live game streaming is tremendously popular, and recent reports indicate
that such platforms impose high traffic volume, leading to degraded user
experience. In this thesis, we propose a Segment-of-Interest (SoI) driven
platform, so as to optimize live game streaming. Our platform uses various
features collected from streamers and viewers combined with sophisticate algorithms
empowered by stat-of-the-art machine learning models to determine
if the current segments of gameplays attract viewers. Upon determining the
importance of individual segments, the limited bandwidth is allocated to the
interested viewers in a Rate-Distortion (R-D) optimized manner, where the
levels of segment importance are used as weights of game streaming quality.
The underlaying intuition is: viewer experience is degraded only when the
game streaming degradation is noticed by viewers. Evaluation results using
real world traces shows that our SoI detecting algorithms can correctly detect
SoI with up to 0.96 F-measure in classifier variant, and up to 0.87 R-squared
score in regressor variant using 10-fold evaluation. Simulation results show
the benefits of our proposed resource allocation solution: (i) it improves viewing
quality by up to 5 dB, (ii) it saves bandwidth by up to 50 Gbps, and (iii)
it efficiently performs resource allocation and scales to many viewers. Our
presented testbed is opensource and can be leveraged by researchers and engineers
to further improve live game streaming platforms.

Acknowledgments i
致謝ii
中文摘要iii
Abstract iv
1 Introduction 1
2 Proposed Architecture 5
3 Research Problems 7
3.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 SoI Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Resource Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4 SoI Detector 11
4.1 Solution Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.3 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.4 Optimal Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.5 SoI Detecting Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5 Resource Allocator 26
5.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.2 Proposed Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.4 Leveraging Features From Viewer . . . . . . . . . . . . . . . . . . . . . 29
6 An Opensource Testbed 32
7 Evaluations 40
7.1 SoI Detector Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.1.1 Evaluation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.1.2 Results From De,R And De,C . . . . . . . . . . . . . . . . . . . . 40
7.1.3 Results From SoI Simulator . . . . . . . . . . . . . . . . . . . . 41
7.2 Resource Allocator Evaluation . . . . . . . . . . . . . . . . . . . . . . . 42
7.2.1 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 42
7.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
8 Related Work 45
8.1 General Live Gaming Streaming Related Research . . . . . . . . . . . . 45
8.2 Large Scale Transcoding . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.3 Video Summarization, Highlight Detection and ROI . . . . . . . . . . . . 46
9 Conclusion and FutureWork 47
Bibliography 49

[1] H. Ahmadi, S. Khoshnood, M. Hashemi, and S. Shirmohammadi. Efficient bitrate
reduction using a game attention model in cloud gaming. In Proc. of IEEE International
Symposium on Haptic Audio Visual Environments and Games (HAVE’13),
Istanbul, Turkey, 2013.
[2] Amazon buys twitch for 970 million in cash. http://www.
businessinsider.com/amazon-buys-twitch-2014-8/.
[3] Amd API homepage. http://tinyurl.com/AMD-ADL.
[4] R. Aparico-Pardo, K. Pires, A. Blanc, and G. Simon. Transcoding live adaptive
video streams at a massive scale in the cloud. In Proc. of ACM Multimedia Systems
Conference (MMSys’15), Portland, Oregon, 2015.
[5] T. F. Chiang, H. Hong, and C. Hsu. Segment-of-Interest driven live game streaming:
Saving bandwidth without degrading experience. In Proc. of IEEE International
Workshop on Network and Systems Support for Games (NetGames’15), Zagreb,
Croatia, 2015.
[6] W. Chu and Y. Chou. Event detection and highlight detection of broadcasted game
videos. In Proc. of ACM Workshop on Computational Models of Social Interactions:
Human-Computer-Media Communication (HCMC’15), Brisbane, Australia, 2015.
[7] J. Deng, F. Cuadrado, G. Tyson, and S. Uhlig. Behind the game: Exploring the
twitch streaming platform. In Proc. of IEEE International Workshop on Network
and Systems Support for Games (NetGames’15), Zagreb, Croatia, 2015.
[8] FFmpeg homepage. https://ffmpeg.org.
[10] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning:
Data Mining, Inference, and Prediction. Springer, 2th edition, 2001.
[11] HTTP live streaming. http://tinyurl.com/HLS-draft.
[12] M. Hossain, M. Muhammad, B. Song, M. Hassan, A. Alelaiwi, and A. Alamri. Audio–
visual emotion-aware cloud gaming framework. IEEE Transaction on Circuits
and Systems for Video Technology, 25(12):2105–2118, 2015.
[13] M. Kaytoue, A. Silva, L. Cerf, W. Meira, and C. Ra¨ıssi. Watch me playing, I am a
professional: a first study on video game live streaming. In Proc. of ACM Workshop
on Mining Social Network Dynamics (MSND’12), Lyon, France, 2012.
[14] Y. Liu, Z. Li, and Y. Soh. Region-of-Interest based resource allocation for conversational
video communication of H.264/AVC. IEEE Transactions on Circuits and
Systems for Video Technology, 18(1):134–139, 2008.
[15] Y.Ma, L. Lu, H. Zhang, andM. Li. A user attention model for video summarization.
In Proc. of ACM International Conference on Multimedia (MM’02), Juan Les Pins,
France, 2002.
[16] M.Margel. Twitch plays pokemon: An analysis of social dynamics in crowdsourced
games. Technical report, University of Toronto, 2014.
[17] A. Mavlankar, P. Agrawal, D. Pang, S. Halawa, N. Cheung, and B. Girod. An
interactive region-of-interest video streaming system for online lecture viewing. In
Proc. of Packet Video Workshop (PV’10), Hong Kong, China, 2010.
[18] G. Nascimento, M. Riberio, L. Cerf, N. Ces´ario, M. Kaytoue, C. Ra¨ıssi, T. Vasconcelos,
and W. Meira. Modeling and analyzing the video game live-streaming
community. In Proc. of Latin American Web Congress (LA-WEB’14), Ouro Preto,
Brazil, 2014.
[19] NGINX homepage. http://nginx.org.
[20] Novabench. https://novabench.com/.
[21] Nvidia API homepage. https://developer.nvidia.com/nvapi.
[22] Open Broadcast Software. https://obsproject.com/.
[24] K. Pires and G. Simon. Youtube Live and Twitch: A tour of user-generated live
streaming systems. In Proc. of ACM Multimedia Systems Conference (MMSys’15),
Portland, Oregon, 2015.
[25] Adobe’s Real TimeMessaging Protocol. http://tinyurl.com/rtmp-spec.
[26] Scikit-learn. http://scikit-learn.org.
[27] R. Shea, D. Fu, and J. Liu. Towards bridging online game playing and live broadcasting:
design and optimization. In Proc. of ACM Workshop on Network and Operating
Systems Support for Digital Audio and Video (NOSSDAV’15), Portland, Oregon,
2015.
[28] SMPlayer homepage. http://smplayer.sourceforge.net/.
[29] I. Sodagar. The MPEG-DASH standard for multimedia streaming over the internet.
IEEE Multimedia, 18(4):62–67, 2011.
[30] G. Sullivan and T. Wiegand. Rate-distortion optimization for video compression.
IEEE Signal Processing, 20(6):74–90, 1998.
[31] Science: Surfing the 4th largest stream of
data. http://blog.twitch.tv/2015/05/
science-surfing-the-4th-largest-stream-of-data/.
[32] BOOM. More transcode servers. http://blog.twitch.tv/2015/07/
boom-more-transcode-servers/.
[33] Twitch retrospective 2013. http://www.twitch.tv/year/2013.
[34] Twitch retrospective 2014. http://www.twitch.tv/year/2014.
[35] Twitch retrospective 2015. https://www.twitch.tv/year/2015.
[36] Windows performance counter website. http://tinyurl.com/
Window-PDH.
[37] Xgboost. http://github.com/dmlc/xgboost.
[38] YouTube gaming website. https://gaming.youtube.com/.
[40] C. Zhang, J. Liu, and H. Wang. Towards hybrid cloud-assisted crowdsourced live
streaming: measurement and analysis. In Proc. of ACM Workshop on Networks and
Operating Systems Support for Digital Audio and Video (NOSSDAV’16), Klagenfurt
am W¨orthersee, Austria, 2016.
[41] Y. Zheng, D.Wu, Y. Ke, C. Yang, M. Chen, and G. Zhang. Online cloud transcoding
and distribution for crowdsourced live game video streaming. IEEE Transaction on
Circuits and Systems for Video Technology, PP(99):1–1, 2016.
[42] X. Zhu, E. Setton, and B. Girod. Congestion-distortion optimized video transmission
over ad hoc networks. Signal Processing: Image Communication, 20(8):773–783,
2005.