Packet loss visibility and packet prioritization in digital videos

Traditional approaches to video quality assume that all packet losses affect quality equally. In reality, different packet losses have different visual impact and not all packet losses are visible to the average human viewer. The problem of evaluating video quality given packet losses is quite challenging due to the varying impact of packet losses. As a first step towards developing a quality metric for video affected by packet losses, we address the problem of predicting packet loss visibility. Visibility of a packet loss refers to the visibility of artifacts in the video caused by that packet loss.; We consider the problem of predicting packet loss visibility as a regression as well as a classification problem. In the regression problem, the goal is to predict the probability that the packet loss causes a visible artifact. In the classification problem, the goal is to classify each packet loss as visible (invisible) if a visible artifact occurred (did not occur) in the video due to that packet loss. A subjective test is conducted to gain ground truth on visibility of 1080 individual packet losses in MPEG-2 videos. We explore the usefulness of various factors in predicting visibility, which are extracted using measurements from either the entire encoded video, the decoded video pixels, or just the received lossy bitstream. We model the probability of visibility of a packet loss using a Generalized Linear Model (GLM). We design classifiers to solve the classification problem using a well-known statistical tool called Classification and Regression Trees (CART).; Video transmission over internet or wireless links is typically characterized by bursty packet losses and not individual packet losses. So we generalize the concept of visibility to multiple losses. A multiple loss is defined as a set of L individual packet losses occurring in close temporal proximity. To obtain ground truth, a new subjective test is conducted using H.264/AVC bitstreams instead of MPEG-2 bitstreams. Further, motion-compensated error concealment (MCEC) is used to conceal the packet losses instead of the zero-motion error concealment (ZMEC) which is used earlier. Because of these differences, 2160 individual packet losses are also introduced along with 3240 multiple losses in this subjective test and the visibility of both types of losses is modeled. We introduce new factors that are useful in predicting visibility. The relative importance of these factors is ascertained through statistical modeling. The effect of different factors on packet loss visibility is also analyzed. A new model framework is introduced for predicting the visibility of multiple packet losses and its performance is demonstrated on dual losses (two packet losses occurring together).; One of the applications for the knowledge of packet loss visibility is packet prioritization. We demonstrate the effectiveness of our visibility model for this application. We consider a transmission scenario where packets are dropped at a congested node in the network. A new packet prioritization method is proposed that assigns a priority level to each packet using our visibility model. We show that a priority-based packet drop policy outperforms a conventional DropTail policy in terms of received video quality.