direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Spatial and Temporal Feature Extraction for Gaming Video Quality Assessment

LOCATION:  TEL, Room Auditorium 3 (20th floor), Ernst-Reuter-Platz 7, 10587 Berlin  

 

Date/Time: 11.11.2019, 14:15-15:00 

SPEAKER: Kumar Awanish (TU Berlin)

Abstract: Gaming video streaming has recently received lots of attention with two paradigms of gaming services, both interactive and passive. The concept of interactive application streaming or cloud gaming (CG) is to render a video game in the cloud and stream the game scenes as a video to players and receiving back the input commands over a broadband network whereas, in passive streaming services like Twitch.tv player gameplay is made available to other users. Both CG and passive video gaming streaming have rapidly expanded their market among gamers and drawn a lot of attention from researchers and businesses. In order to ensure user satisfaction of services, it is important to keep the end-user gaming quality of experience (QoE) high. Video quality is one of the important quality features that influence gaming QoE. In contrast to traditional video content, gaming content has special characteristics such as an extremely high motion for some games, special motion patterns, synthetic content and repetitive content, which creates new challenges during the video quality assessment. Due to this inherent nature and different requirements of gaming based video streaming services, there is a need for application-specific lightweight, No-Reference(NR)  gaming video quality prediction models.

In this thesis, we present a machine learning-based lightweight No-Reference (NR) metric, named NR-GCM, for the quality assessment of gaming videos. The proposed model consists of two stages in the design process of the model and uses spatial and temporal features for the model development. The first stage involves the selection of best spatial features on the frame-level then train a regression model to predict the quality of each frame and save the trained frame-level model to use in stage two. In stage two, the learned  representation from stage one is sent to another regression model along with the best selected temporal features on video-level. The final output is a NR metric NR-GCM, which does video quality assessment (VQA). We evaluate the performance of the proposed model  on different gaming video datasets and show that the proposed models outperform the current state-of-the-art lightweight No-Reference metrics. In addition, the performance of the proposed model is comparable with the best known Full-Reference(FR) metric. 

Zusatzinformationen / Extras

Direktzugang

Schnellnavigation zur Seite über Nummerneingabe