Pravin Rajamoney

CSE 581 – Network Technology

 

Topic:  Video Modeling

 

            The presentation was based on the 4 papers attached to this document.  The papers cover the idea of video modeling and the advantages it can provide.  The basic MPEG video compression concept is to do much of the work (compression) up-front and the decoder doesn’t need to work too hard to decompress it and display the content.

 

            The advantage Constant Bit Rate (CBR) video has over VBR is that modeling-using CBR is easier.  The CBR model is a fluid flow model.  Buffer management on the encoding and decoding is easier to manage.  This helps makes data flow on the network much easier to control and maintain.  Unfortunately, it is not bandwidth efficient.  Maximum bandwidth of the stream must be always allocated for that stream.

 

            The Variable Bit Rate (VBR) stream on the other hand is better on network bandwidth efficiency.  Data is variable or bursty which is in common with Ethernet data on a network.  The disadvantage of VBR is that it requires careful buffer management and data control.  Modeling VBR traffic is not trivial since traffic is dependent on the video content.

 

            Previous modeling of the video only took a subset of a video to model a video stream.  It has been found that data taken from Short-Range Dependent models were not sufficient to correctly model VBR video.  This model used about 10ms of data or less then 200 frames to model a VBR video stream.  The Long-Range Dependent or the “Hurst Parameter” is a better video model.  The definition for LRD is “Observation of an empirical record being significantly correlated to observation that are far removed in time”.  Hence the autocorrelation for data although very small still needs to be evaluated when modeling a VBR stream

 

            According to the Hurst Parameter, the VBR stream is dependent on the contents.  For example, a sports event will have a higher LRD value then a talk show.  And so when modeling video for a VBR stream that has “talking heads”, the average bit rate will be much lower.  Buffer management will have to correspond to incoming data and depending on the content, adjust accordingly. 

 

            The class discussion started on what a negative autocorrelation value meant on one of the figures in the papers? Another question was why is buffering an issue if memory is cheap and large in most machines nowadays?   Buffering will help in jitter but would cause latency on a streaming video.  This may or may not be a problem on some applications as long as there is enough for the time for the decoder to buffer the stream and decode it.