We decided to leave out fields and keep video progressive. There are the I, P, and B frames, which need to be sequentially numbered in the original and remembered in partitioned versions. A bit pattern in the descriptor would work, and could be kept small using a fourier analysis plus run-length encoded supplemental bits instead of the complete bit pattern. Although even for a two hour 30fps video, the bit pattern remains pretty managable.

Which DCT coefficients are kept also needs a bit pattern. Since the idea is to let different versions be put together to form a higher quality version, the two versions need to pick different coefficients to keep. The DCT coefficient bit pattern indicates which ones are kept.

For the motion vectors, we decided to not throw them out since reconstructed video probably wouldn't be very valuable without them. However, we require no residual motion vectors, and since portions of the video plane may be cropped out, sometimes external block data will need to be carried over since motion vectors are differentially encoded.

There is also the q-scale, which changes on a block by block basis. Choosing which block is better requires going through both versions of a video and picking those with the lower q-scale. However, we want to be able to determine if there is any real value to be gained by this pass without actually doing the pass. One way would be to generate a hash or checksum of the q-scale values to identify if two versions with otherwise identical descriptors may result in a higher quality reconstructed version. But this still requires you to go through the entire video to find what might simply be a single possible point of improvement. So something better would be much more useful.

Post a comment