Re: h.264 intra prediction

Malcolm Taylor wrote:

I am assuming that you are referring to motion compensation, and in this case I can answer your question. 
Rather the opposite. Intra prediction is a prediction used in I type
macroblocks where no reference frames samples are used.

For motion compensation, the receiver must know what block was used for the prediction in order to reconstruct the image. In JBIG, the predicting pixels are always the same, and so no side information needs to be transferred. However in h.264 (as in most modern codecs), the block which is used for prediction is identified by a motion vector. This vector must be sent to the decoder, and it takes up space in the output.
If you were to predict individual pixels, then a vector would need to be sent for each pixel. That makes for a very large amount of side information, so much so that it is not useful if your aim is compression! :) 
This is not obvious. H.264 uses predictive coding for motion vectors so
only difference to prediction is stored. Furthermore arithmetic coding
is available hence storing a lot of [0,0] vector differences would not
take much bitrate. The real problem here is encoding/decoding time.


The block sizes used form a good trade off between the space required to write the motion vectors, and the accuracy of the prediction.

Malcolm


--
m.
.

Relevant Pages

  • Re: h.264 intra prediction
    ... For motion compensation, the receiver must know what block was used for the prediction in order to reconstruct the image. ... In JBIG, the predicting pixels are always the same, and so no side information needs to be transferred. ... However in h.264, the block which is used for prediction is identified by a motion vector. ...
    (comp.compression)
  • video compression algorithm
    ... The first three channels are for the independent part of the frame. ... The next three channels are for the prediction residuals ... transformations of 2 or 4 pixels. ... motion vectors, and the prediction bitmap, then a transformation of 3 ...
    (comp.compression)
  • Re: Prediction in differencing pixels with the left neighbour.
    ... for example predict odd pixels by the average of the two neighbouring even pixels - thus implementing the high-pass of the 5/3-wavelet - and then update the even pixels from the odd pixels to keep the average identical. ... Another direction would be a non-linear prediction from nearest neighbours, for example the predictor of JPEG-LS (check for Weinberger/Sapiro, and the "LOCO" coder). ...
    (comp.compression)
  • Re: Understanding wavelet implementation
    ... perform a prediction using a predictor based on average of ... left and right pixels. ... *Quantizing* the coefficients and then reconstructing might suppress noise because it suppresses higher frequencies in the image that are often typical for additive noise. ...
    (comp.compression)
  • Talking about prediction
    ... Could you share your prediction matrix which you use for the pixel ... I am desperately searching for a nonlinear predictor for ... grey pixels. ...
    (comp.compression)