One Constant Instead of a Lookup Table Trimming the Arithmetic Inside H.266 Matrix Intra Prediction

This application covers a narrow but heavily used piece of arithmetic in the H.266/Versatile Video Coding standard. Matrix-based Intra Prediction, or MIP, predicts a block of pixels by multiplying neighbouring reference samples through a stored weight matrix, and the result then has to be scaled back down by a right shift. In the version of MIP the specification argues with, the size of that shift changed with both the shape of the block and which MIP mode was chosen, so encoders and decoders had to carry a table of shift values and look one up for every block. The application, filed by OPPO, replaces that table with a single constant.

The Problem

The specification opens with the usual pressure on video coding. High definition and ultra high definition formats keep arriving, H.265/HEVC can no longer keep up with them, and the Joint Video Experts Team responded with H.266/VVC and its reference software, the VVC Test Model. MIP was one of the tools accepted into that standard. It sorts intra coded blocks into categories and gives each category a different number of MIP modes, each mode backed by its own weight matrix.

The complaint is arithmetic rather than conceptual. MIP builds an integer matrix by scaling the weights up, multiplies it against a reduced set of reference samples, and then shifts the product back to the right to recover a usable prediction, with an optional rounding term added before the shift.

The difficulty is what governs the size of that shift. In the MIP version the specification describes, it is a function of two things at once: the block size category, and the index number of the particular MIP mode in use. The specification reproduces the resulting table, three rows of block size categories against eighteen columns of mode indices, and almost every entry is the number 6. A handful are 5 and a handful are 7. That table has to sit in memory as a lookup table on the encoder side, and every block costs a lookup. For a value that is overwhelmingly the same number, the specification treats this as an unreasonable price paid on both memory and calculation complexity.

What This Invention Does

Claim 1 is an encoder side method for colour component prediction. It determines the prediction parameters of the current block, being a prediction mode parameter and a size parameter. When the prediction mode parameter says MIP is to be used, it determines three items, the MIP weight matrix, the shift factor and the MIP input sample matrix, and produces the intra prediction value from those three. The operative limitation is the next phrase: the shift factor of the current block comprises a constant shift factor. The mode dependent table is gone.

The rest of claim 1 pins down where the weight matrix comes from. The MIP weight matrix is determined at least according to a block size index value, and that index is derived from the size parameter by an explicit three way test written into the claim. If both the width and the height of the block are 4, the index is 0. If both are 8, or if either the width or the height is 4, the index is 1. If neither condition holds, the index is 2. Claim 2 fixes the constant at 6, the value that dominates the discarded table, though the specification also lists 5 and 7 as workable choices.

That index is not only a label for the weight matrix. It also selects how many boundary reference samples are taken from each side and how large the matrix multiplication output is: two samples per side and a 4×4 prediction block at index 0, four per side and a 4×4 block at index 1, four per side and an 8×8 block at index 2. Making the shift constant and keeping the weight matrix keyed to that same index means a single derived number drives the whole configuration, with no second dimension of mode index to carry.

Key Features

• A constant shift factor. The right shift applied after MIP matrix multiplication is fixed rather than derived per block and per mode, which is the limitation claim 1 turns on.
• A value of 6. Claim 2 sets the constant at 6, the value that already filled most cells of the table it replaces, so the result changes for only a minority of block and mode combinations.
• A three way block size index written into the claim. The 4×4 case, the 8×8 or contains a side of 4 case, and everything else, mapped to index values 0, 1 and 2.
• The weight matrix keyed to block size alone. The MIP weight matrix is determined at least according to that index value, not according to a joint function of size and mode.
• A smaller lookup table. Where a table is still used, removing the mode index dimension shrinks what has to be held in memory, a saving the specification names alongside reduced calculation.
• Matched encoder and decoder methods. The same logic is claimed from the decoder side, where the prediction parameters are read out of the bitstream instead of chosen, so both ends derive the same shift without it being signalled.
• Claimed as apparatus and storage medium. Beyond the method, the filing covers an encoder and a decoder with memory and processor, and a computer storage medium holding the program.

Who Is Behind It

The applicant is Guangdong OPPO Mobile Telecommunications Corp., Ltd., the Dongguan based consumer electronics manufacturer that sells phones under the OPPO brand. Companies at that scale file heavily on video coding because every device they ship contains a decoder, and standard essential patents in a codec are worth more as a licensing position than as a product feature.

Three inventors are named: Junyan Huo, Yanzhuo Ma and Wei Zhang. The same group appears on a long run of OPPO video coding filings from the same period, most of them tracking proposals made into the JVET process while H.266 was being finalised.

This is a divisional of Australian application 2020310350, filed on 23 March 2021, which claims priority from three United States provisional applications: 62/872,488 filed on 10 July 2019, and 62/872,830 and 62/873,170 both filed on 11 July 2019. Those dates place the work in the middle of 2019, while MIP was still being refined inside the draft standard.

Why It Matters

Coding standards are settled by argument over exactly this kind of detail. A tool is adopted, then successive meetings strip out cost that does not buy compression, and a proposal that removes a lookup table without measurably hurting quality is the sort of thing that gets folded into the text with little fanfare. MIP went through several rounds of that treatment during 2019, before the final standard was published by the ITU in 2020 and mirrored in the reference software.

The commercial logic of a divisional filed in 2026 for work done in 2019 is coverage rather than novelty. If a patent family reads on a mandatory part of a widely deployed codec, its value comes from the breadth of the claims that survive examination, and filing successive divisionals lets an applicant keep prosecuting alternative claim shapes long after the parent has settled. Australia’s divisional practice makes that a routine tactic.

H.266 decoders are now shipping in hardware, and every one of them implements MIP. Arithmetic simplifications are worth pursuing in that setting because they multiply across every intra coded block in every frame, and because the cost lands in fixed function silicon where memory for a lookup table is an area budget rather than an abstraction.

Related Concepts

• Intra-frame coding – the class of prediction MIP belongs to, using only samples from within the same picture.
• Matrix multiplication – the operation at the centre of MIP, and the reason a scaling shift is needed at all.
• Arithmetic shift – the right shift operation whose size this application makes constant.
• Versatile Video Coding – the H.266 standard that adopted MIP as an intra prediction tool.
• High Efficiency Video Coding – the predecessor standard the specification says can no longer meet demand.
• Chroma subsampling – background to why colour components are predicted and coded separately from luma.

---

AU 2026202130 was published in the Australian Official Journal of Patents on 9 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.