On the accuracy of short-term quality models for long-term quality prediction

With video services such as HTTP-based adaptive streaming, network congestion may result in quality fluctuations over several minutes. There is therefore a need for estimating the quality of long audiovisual sequences. This can be achieved by using short-term audiovisual quality models, which output quality scores for short periods of time, for instance 10 s. Temporal pooling such as averaging is typically applied on the short-term quality estimates for providing a quality score for a longer time period, for instance three minutes. With this modeling strategy, the performance of the overall quality model can be increased by improving both the short-term quality model and the temporal pooling strategy. However, depending on the temporal pooling strategy, and possibly the targeted test data obtained for long sequences, a small improvement of the short-term quality model may eventually not have any significant impact on the long-term quality estimates. This paper investigates this aspect by comparing the performance results of the combination of six short-term quality models with six different pooling strategies. Results show that the performance of well performing short term models is a good indicator of the performance of the long-term quality models, independently of the pooling strategy.