Motion Prediction In Video Coding

What is this patent about?

’354 is related to the field of video coding and, more specifically, to improving the efficiency of motion-compensated prediction in video codecs. Modern video codecs use motion compensation to predict pixel values in a frame based on previously encoded frames. This involves finding similar blocks in reference frames and encoding only the difference, or residual, between the predicted and actual pixel values. A key challenge is accurately predicting pixel motion, especially when blocks move by fractional pixel amounts.

The underlying idea behind ’354 is to reduce rounding errors that accumulate during bi-directional or multi-directional prediction. Instead of rounding prediction signals to the original pixel precision after each prediction step, the patent proposes maintaining the prediction signals at a higher precision throughout the prediction calculation. Only after combining multiple prediction signals is the final result rounded down to the original precision. This approach minimizes the impact of rounding errors on the final prediction.

The claims of ’354 focus on a method, apparatus, and computer program product for decoding or encoding a block of pixels. The core steps involve determining reference pixel locations in two or more reference blocks, generating predictions from each reference block at a higher precision than the original pixel values, combining these high-precision predictions, and then reducing the precision of the combined prediction by bit-shifting.

In practice, the invention involves using motion vectors to locate corresponding blocks in reference frames. Interpolation filters are then applied to obtain prediction signals at a higher bit depth (e.g., 16 bits instead of 8 bits). These high-precision signals are combined, and finally, the combined signal is shifted to reduce the bit depth back to the original pixel precision. This can be implemented in both the encoder and decoder to ensure consistent prediction.

This approach differs from prior methods that perform rounding after each individual prediction, which can lead to accumulated errors. By maintaining higher precision throughout the calculation, the invention avoids the need to signal rounding offsets or use different rounding methods for different frames, simplifying the codec and potentially improving coding efficiency. The bit-shifting operation provides a computationally efficient way to reduce the precision of the final combined prediction.