An objective and subjective quality assessment for passive gaming video streaming

Gaming video streaming has become increasingly popular in recent times. Along with the rise and popularity of cloud gaming services and e-sports, passive gaming video streaming services such as Twitch.tv, YouTubeGaming, etc. where viewers watch the gameplay of other gamers, have seen increasing acceptance. Twitch.tv alone has over 2.2 million monthly streamers and 15 million daily active users with almost a million average concurrent users, making Twitch.tv the 4th biggest internet traffic generator, just after Netflix, YouTube and Apple. Despite the increasing importance and popularity of such live gaming video streaming services, they have until recently not caught the attention of the quality assessment research community. For the continued success of such services, it is imperative to maintain and satisfy the end user Quality of Experience (QoE), which can be measured using various Video Quality Assessment (VQA) methods. Gaming videos are synthetic and artificial in nature and have different streaming requirements as compared to traditional non-gaming content. While there exist a lot of subjective and objective studies in the field of quality assessment of Video-on-demand (VOD) streaming services, such as Netflix and YouTube, along with the design of many VQA metrics, no work has been done previously towards quality assessment of live passive gaming video streaming applications. The research work in this thesis tries to address this gap by using various subjective and objective quality assessment studies. A codec comparison using the three most popular and widely used compression standards is performed to determine their compression efficiency. Furthermore, a subjective and objective comparative study is carried out to find out the difference between gaming and non-gaming videos in terms of the trade-off between quality and data-rate after compression. This is followed by the creation of an open source gaming video dataset, which is then used for a performance evaluation study of the eight most popular VQA metrics. Different temporal pooling strategies and content based classification approaches are evaluated to assess their effect on the VQA metrics. Finally, due to the low performance of existing No-Reference (NR) VQA metrics on gaming video content, two machine learning based NR models are designed using NR features and existing NR metrics, which are shown to outperform existing NR metrics while performing on par with state-of-the-art Full-Reference (FR) VQA metrics