MPEG-2 Custom INTER Matrix: Linear Local Variation Approach via Binary Convergence

[manuel]

Hi everyone,

I am designing a custom quantization matrix for low-bitrate MPEG-2 encoding, analyzed through the mathematical binary convergence to absolute zero (using 255 as the target numerator).

For my INTRA matrix, I built a perfect diagonal shell structure using Even Hybrids to maximize stability and prevent flickering.

Now, for the INTER matrix, I am experimenting with a strict "Linear Local Variation" constraint. My goal is to reach a maximum 2x (+100%) compression step in the last high-frequency cell.

To achieve a perfectly smooth rampa, I am distributing the compression load proportionally across the 63 steps using a fixed cell-to-cell increment formula:
(100 - 0) / 63 = ~1.5873% for each local variation step based on the worst-case scenario of integer quotients.

What are your thoughts on using this constant percentage progression for smoothing out motion vectors in HCEnc?

Thanks,
Manuel

[rwill]

There are mainly 2 options:

1. Have different INTRA and INTER matrices and add strong dithering to the input to gain stability and to prevent flicker. Due to the dithering it will flicker anyway.

2. Have the same matrix for INTRA and INTER. This will make things stable and there will be almost no flicker.