Viewing impaired video transmissions from a modeling perspective

Abstract

The provisioning of broadband access for mass market enables multimedia content to be extensively viewed and exchanged over the Internet. Consequently, Quality of Experience (QoE) aspects, describing the service quality perceived by the user, become vital factors in ensuring customer satisfaction in today’s networks. In the case of streaming video, enforcing QoE is a complex research topic [4] involving various aspects such as the characteristics of network traffic, codec functions, and measures of user perception of video quality. For instance, the impact of packet loss on the user’s perception of real-time video can be investigated by using measured packet loss traces or generated by analytical models. These traces are further used to impair video sequences at different packet loss rates, whereas visual quality properties are then derived by decoding the produced sequences. Both measurement and model based generators for loss traces have advantages and disadvantages. A loss trace obtained from measurements accurately reflects the state of the network. On the other hand, fitting statistical properties observed in measurements to predefined loss models allows producing multiple loss traces with various properties, such as packet loss rates and loss burst length. This flexibility permits studying the behavior of video sequences impaired under different error conditions. A disadvantage of model generated loss traces is that the statistical properties may not adequately replicate those of the measured trace, as they are likely to be biased by model limitations. Consequently, analytical models are needed to closely replicate the loss patterns observed in measurements. Finite-state Markov chains are a particular class of models for generating packet loss traces. In general, these models are widely used to characterize error processes in telecommunication systems and to evaluate the performance of coding or other measures for error resilience [5, 6, 9]. A large body of work focuses on replicating loss burst and gap length distributions which are relevant for evaluating the performance of FEC algorithms. In the case of encoded video, however, errors over larger time-scales than burst and gaps may cause various visual effects, e.g., distorted areas of an image. Depending on the bitrate, time-scales in video can range over multiple orders of magnitude from microseconds in the case of the transmission of a single macroblock—a small part of a still image—up to seconds for the transmission of an entire group of pictures. Moreover, with the advent of the latest video encoding standard H.264 [3], parameter sets describing a set of pictures with similar decoding properties can introduce even larger time-scales. While losing a single macroblock can have a minor impact by affecting only a small fraction of the image, losing a parameter set can have a severe effect on the decoding process. This consideration suggests to focus on the loss process over multiple time-scales rather than on matching burst and gap length distributions. To address this observation, this paper uses an M -state Markov chain to generate loss traces for impairing video transmissions. To this end, Section 2 provides a general result for M -state Markov chains to describe the distribution of the packet losses over multiple time-scales using secondorder statistics. By using moment matching, 2-state Markov models are fitted to a wireless Digital Video Broadcasting (DVB-H) loss trace. Then Section 3 presents numerical results which illustrate that the visual impairment patterns depend on the chosen modeling technique. For future work we plan to investigate relevant time-scales for the fitting process. Moreover, we plan to evaluate whether increasing the number of states of the underlying Markov chain results in any significant gain. This paper is an extension of [1] where we mainly focused on fitting 2-state Markov models to DVB-H and backbone packet loss traces using secondorder statistics.