On September 16th, we'll be going to the NCNI meeting to present a few different proposals as to how we can go about developing a tool to predict video quality based on network measurements. So between now and then, my job is to find perceptual image quality models which I think work well and also apply to H.263 and MPEG-2. Given these models, I need to figure out what sort of network measurements will let us predict the final video quality.

I've already got the Sarnoff JND model which I do think works well. In particular, this model applies to DCT images in general, rather than video streams. So it should work for both H.263 and MPEG-2. We should be able to figure out the probabilities for different levels of image reconstruction quality given some network characteristics and traffic statistics. And then applying the JND model to the reconstructed images will give us probabilities for different video qualities under those network conditions.

So, other than looking for some additional perceptual image quality models, I also need to look up H.263 implementations to see which are the most commonly used features. Different features will affect the traffic characteristics and image reconstruction. There might be some papers which deal specifically with H.263.

One thing which will come later on is figuring out how we can generate traffic with video characteristics to gather link statistics and measurements. Ketan suggested a Java applet which generated the traffic, but we both have doubts as to how well that would work. A Java applet is not going to be very reliable for precise time measurements. But it would be easier to implement than Ketan's other suggestion. If we point a web browser at our own server which is modified to send back data mimicking video traffic, we can make measurements at the TCP/IP layer. I think this approach is both more accurate and more elegant from a user's point of view.

Post a comment