Skip to Main Content
Current objective video quality metrics typically estimate video quality for short video sequences (10 to 15 sec) of constant quality. However, customers of video services usually watch longer sequences of videos which are more and more delivered via adaptive streaming methods such as HTTP adaptive streaming (HAS). A viewing session in such a setting contains several different video qualities over time. In order to express this in an overall score for the whole viewing session, several temporal pooling methods have been proposed in the related work. Within this paper, we set out to compare the performance of different temporal pooling methods for the prediction of Quality of Experience (QoE) for HTTP video streams with varying qualities. We perform this comparison based on ground truth rating data gathered in a crowdsourcing study in the context of the NGMN P-SERQU project. As input data for the models, we use objective video quality metrics such as PSNR, SSIM but also very basic inputs such as the bitrate of the clips only. Our results show that certain pooling methods perform clearly better than others. These results can help in identifying well performing temporal pooling methods in the context of HAS.